Datasets:

MarkrAI

/

AutoRAG-evaluation-2024-LLM-paper-v1

like 8

1b342903-605b-46d0-bf8c-44525c0d0212

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.3. Text To Video | | All | Project journey montage | Implement generative AI to compile a video | | montage showcasing the entire project journey, | | | | combining textual descriptions and visual | | | | elements for a comprehensive overview. | | |

2657876e-07ce-4322-90f1-39e046b81864

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.4. Image To Text Applying AI in converting images into text significantly improves construction procedures at all stages. Converting images into text descriptions also has valuable applications throughout construction project phases. **Table 15** summarizes potential use cases validated by the expert discussion. Pre-construction opportunities include extracting measurements, boundaries, and other information from land surveys and blueprints. During construction, daily site photos could be analyzed to generate progress reports. Images of materials and equipment could develop real-time quality and inventory assessments via generative AI [105]. Figure 7 displays an image description generated by Gemini Pro as part of a daily visual report. Upon close inspection, the model accurately captures fine-grained details in the image, including identifying the specific brand and model of the construction equipment. This demonstrates Gemini Pro's capability to produce descriptive text summarizing critical visual information [95]. Post-construction applications involve extracting as-built details from archival photos and making warranty documentation from damaged images. Cross-cutting use cases include automating visual inspection reports across phases. With proper training, image captioning techniques can translate construction graphics and photos into structured textual information. This eliminates tedious manual efforts to log and convey visual observations. Models such as GPT-4 can analyze everyday images and accurately describe prominent objects, actions, and scenery [45]. | Potential opportunity | Description | Project phase | |-------------------------------------------------|-----------------------------|------------------------------------------------| | Pre-construction | Land survey data extraction | Analyze land survey images and extract textual | | data, such as measurements, topographical | | | | details, and boundary information.

48de09e1-7108-4c59-b511-e806cbd06447

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.4. Image To Text | Land survey data extraction | Analyze land survey images and extract textual | | data, such as measurements, topographical | | | | details, and boundary information. | | | | Pre-construction | 3D model specification | Analyze 3D architectural models and | | automatically generate detailed textual | | | | specifications of materials, components, | | | | dimensions, etc. |

83106f2e-48dc-4d61-a551-6fc814e2c66f

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.4. Image To Text | | | | dimensions, etc. | | | | Pre-construction | Blueprint digitization | Automatically convert scanned paper | | blueprints and hand-drawn sketches into | | | | structured digital representations. | | | | Construction | Daily progress image |

5cfb6c0a-be20-4896-99f5-6f03b4a2e503

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.4. Image To Text | | | Construction | Daily progress image | | | analysis | | | | Analyze daily progress images from | | | | construction sites and generate textual reports | | | | summarizing the progress, challenges, and |

0bc4e41f-5791-493d-a330-bee68f9a11b9

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.4. Image To Text | | construction sites and generate textual reports | | | | summarizing the progress, challenges, and | | | | achievements. | | | | Construction | Material quality assessment | Examine images of construction materials and | | generate textual assessments regarding quality, | | | | potential issues, and compliance. | |

caa88642-cd2b-40dd-b705-63379c65b8b8

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.4. Image To Text | | | potential issues, and compliance. | | | | Construction | Inventory management | Use AI techniques to automatically catalog on- | | site equipment, materials, tools, etc., from | | | | images and videos into searchable inventory | | | | databases. | |

3d377e9f-ba2b-4e0e-a24f-f317fb91e560

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.4. Image To Text | | databases. | | | | Post-construction | As-built documentation text | | | extraction | | | | Analyze images of as-built documentation and | | | | extract textual information, facilitating the | |

d1a9e601-809a-4419-af4d-1ec147f6b2e2

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.4. Image To Text | | extract textual information, facilitating the | | | | creation of detailed post-construction reports. | | | | Post-construction | Warranty claim | | | documentation | | | | Explore images of construction components | |

528011cb-8429-434e-a5b4-67906c318ac2

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.4. Image To Text | | | Explore images of construction components | | | | and generate textual documentation for | | | | warranty claims, specifying issues and relevant | | | | details. | | | | All | Visual inspection reports | Examine images from visual inspections and

7bb2bb06-320a-4474-8504-77d8fb49b29f

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.4. Image To Text | | | All | Visual inspection reports | Examine images from visual inspections and | | generate textual reports, providing detailed | | | | information on observed conditions and | | | | recommendations. | | |

6cf2ba54-08c8-4606-bb79-c80911609463

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.5. Image To Image The image-to-image capabilities of generative AI are crucial in construction as they facilitate design adjustments, on-site issue solutions, and the visualization of future upgrades. **Table 16** details potential applications of image-image models in the construction industry. Preconstruction use cases include adapting architectural visualizations into different desired art styles and refining scanned maps into clear site plans. During construction, input architectural plans and sketches could be auto-modified to match ongoing changes on-site [28]. Material texture libraries could help generate realistic composite renderings from sample images. Postconstruction applications involve visualizing the restored building state from damage assessment images and landscape enhancements. Improving image quality and resolution are potential applications across all phases of construction projects. Techniques such as pix2pix GANs demonstrate capabilities to transform input images while preserving essential content structure [99]. By learning alignments between construction image domains during training, models can translate inputs into desired stylistic, structural, or conceptual outputs. This allows the adaptation of visual data into appropriate formats for downstream usage, reducing repetitive manual editing. For instance, rough sketches produced during early design phases can be refined into polished architectural visualizations or engineering schematics. Images captured on-site can be adapted to match design intent, even when physical conditions vary. Continued advances in high-resolution GANs will further expand the potential for image-to-image synthesis to enhance visual media throughout construction projects. | Potential opportunity | Description | Project phase | |---------------------------------------------------|---------------------------|------------------------------------------------| | Pre-construction | Architectural image | | | translation

b5b4effc-6ff7-4d17-a341-7965c11515b6

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.5. Image To Image |------------------------------------------------| | Pre-construction | Architectural image | | | translation | | | | Use generative techniques to adapt | | | | architectural visualizations and renderings | | | | done in one style to different target art styles. | | | | Pre-construction

529f229c-fb75-46a5-85e0-bb9f25d261ef

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.5. Image To Image | | done in one style to different target art styles. | | | | Pre-construction | Site planning refinement | Refine scanned maps and satellite imagery to | | generate clear site/lot diagrams and top-down | | | | site plans for planning. | | | | Pre-construction | Concept generation | Produce variations of initial architectural | | sketches and concept art to explore broader | |

8027f645-a298-467b-9e23-4b4e4d910d9c

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.5. Image To Image | Concept generation | Produce variations of initial architectural | | sketches and concept art to explore broader | | | | design possibilities. | | | | Construction | Updating architectural | | | drawing | | | | Modify architectural drawings and plans by | |

1b51238a-dc5f-4549-8c61-797bd6c39091

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.5. Image To Image | | | | Modify architectural drawings and plans by | | | | incorporating changes made on the | | | | construction site to keep documentation up-to- | | | | date. | | | | Construction

b8168ded-6b8c-40d6-9d63-197b601b7968

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.5. Image To Image | | | | Construction | Material texture matching | Apply generative AI to match the textures of | | construction materials with reference images, | | | | ensuring consistency and quality in the visual | | | | appearance of the constructed elements. | | | | Post-construction | Damage assessment | Process images of damaged building areas and | | generate visualizations showing restored states. |

2729418a-9f04-4e84-98d9-a5593258aa53

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.5. Image To Image | | Post-construction | Damage assessment | Process images of damaged building areas and | | generate visualizations showing restored states. | | | | Post-construction | Landscape transformation | | | visualization | | | | Visualize the transformation of landscapes | | | | based on input images, supporting post- |

7bc7a91a-52d4-4dda-9d92-d51dc916d2b2

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.5. Image To Image | | Visualize the transformation of landscapes | | | | based on input images, supporting post- | | | | construction projects such as garden | | | | enhancements or environmental modifications. | | | | All | Aesthetic enhancement | Improve resolution, lighting, orientation, and | | low-quality construction images across phases. | |

876d20eb-c99e-46a4-94d3-7d96b9d9e8fe

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.5. Image To Image | Aesthetic enhancement | Improve resolution, lighting, orientation, and | | low-quality construction images across phases. | | |

c81655b9-512b-4cc2-a083-5a9e074aa601

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.6. Image To Video Converting static images into dynamic videos opens up impactful possibilities across the construction project lifecycle. **Table 17** summarizes potential applications in this area. During pre-construction, static architectural concept images could be converted into engaging animated walkthroughs and fly-throughs to showcase designs. Aerial site photos could also produce simulated planning and development timelapses [106]. Safety incidents could be recreated on active construction sites based on analysis of images of unsafe conditions to improve hazard awareness through vivid video representations. Timelapse build videos compiled from daily construction photos help visually track project progression. Postconstruction use cases include generating promotional experience videos from facility images and collecting recap documentary videos from archival visuals. State-of-the-art generative video models demonstrate increasing capabilities to animate photo-realistic footage from sparse image inputs [50]. AI systems can extend single images into complete video sequences with convincing continuity and realism by learning to extrapolate motion and physical interactions [107]. Construction visuals contain extensive intrinsic structures that video generation models can leverage to produce meaningful video representations without full frame-by-frame supervision. Converting images into dynamic videos helps improve engagement and understanding compared to static depictions alone. As the coherence and resolution of image-to-video models continue improving, their potential applications in construction for bringing visuals to life will grow. | Potential opportunity | Description | Project phase | |------------------------------------------------------------|---------------------------|--------------------------------------------| | Design concept visualization Generate animated walkthrough | Pre-construction | | | visualizations of architectural concept | |

4b63e4e6-3718-41ca-a520-4c301a32df1a

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.6. Image To Video ruction | | | visualizations of architectural concept | | | | designs from still images. | | | | Pre-construction | Site planning simulation | Produce simulated timelapse videos of site | | planning and layout from aerial photos. | | | | Construction | On-site safety analysis | Assess images of unsafe conditions and | | generate simulated incident recreations for

711d40a0-94c4-44dc-84f2-2dcfe1ed3912

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.6. Image To Video | | Construction | On-site safety analysis | Assess images of unsafe conditions and | | generate simulated incident recreations for | | | | safety analysis. | | | | Construction | Construction progress | Compile timelapse videos of construction | | progress from daily site images. | | | | Post-construction

1b902ea7-119e-4d33-953e-b6c57639fa0d

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.6. Image To Video | Compile timelapse videos of construction | | progress from daily site images. | | | | Post-construction | Operation training videos | Produce equipment maintenance and | | operation training videos from instruction | | | | manual images and diagrams. | | | | Post-construction | Promotional videos | Automatically generate engaging facility | | experience videos from images for |

714c9041-2004-42df-b7fd-4c3715cd62db

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.6. Image To Video | | Post-construction | Promotional videos | Automatically generate engaging facility | | experience videos from images for | | | | leasing/sales. | | | | All | Documentary videos | Compile construction progress, milestones, | | interviews, etc., into documentary-style | | | | recap videos from images.

c2959859-cbbe-4631-b4cc-9787b2244b58

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.6. Image To Video into documentary-style | | | | recap videos from images. | | |

8306b6c8-af39-4673-9842-002868917d82

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.7. Video To Text AI transcription of video-to-text revolutionizes the construction industry by providing comprehensive documentation throughout the construction process. Potential applications of video-to-text conversion are shown in **Table 18.** During pre-construction, generative models could auto-transcribe kickoff meetings and regulatory compliance tutorial videos into concise text records. Safety briefing videos and daily progress meeting discussions on active construction sites could be translated into text summaries for distribution to wider stakeholders [104]. Post-construction commissioning and inspection videos also contain valuable verbal feedback that video-to-text techniques can structure into reports. Across phases, comprehensively transcribing archived project videos into indexed, searchable documentation enables robust retrospective analysis [105]. Models such as VideoCoCa can transcribe technical construction multimedia while filtering out irrelevant background noise [100]. The text outputs synthesize the key details and language from videos without needing to review hours of footage. This allows scaling extraction of vital audio information in rich multimedia that construction teams continuously generate. When deployed with proper data controls, video-to-text AI can unlock new levels of value from archived construction data without demanding extensive manual effort. | Potential opportunity | Description | Project phase | |-----------------------------------------------|----------------------------|----------------------------------------------| | Pre-construction | Meeting minutes generation | Automatically generate minutes from video | | recordings of project kickoff and planning | | | | meetings. |

3a4a8e51-38e0-4322-a5f1-90d0f0341e98

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.7. Video To Text | | | meetings. | | | | Pre-construction | Regulatory compliance | | | briefs | | | | Transcribe spoken content from regulatory | | | | compliance videos, creating textual briefs |

fec173bc-ffa1-4d4d-b6a9-8a75bb7b15ab

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.7. Video To Text | | Transcribe spoken content from regulatory | | | | compliance videos, creating textual briefs | | | | summarizing compliance requirements for | | | | the construction project. | | | | Construction | Safety briefing text | | | summaries |

61c6c6ab-1d71-4007-93f9-076fa6a12fc9

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.7. Video To Text | Safety briefing text | | | summaries | | | | Use generative AI to transcribe safety | | | | briefings in construction videos, generating | | | | concise text summaries for distribution to | | | | construction teams and stakeholders. | |

45e0e472-66f5-4a67-bec5-0cc565475d2a

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.7. Video To Text | | | | construction teams and stakeholders. | | | | Construction | Defect detection | Analyze inspection videos and generate | | written alerts about potential issues for | | | | remediation. | | | | Construction | Daily progress meeting |

7947452b-fc4f-4721-a4d1-5f3767ea9280

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.7. Video To Text | | | Construction | Daily progress meeting | | | transcripts | | | | Transcribe discussions from daily progress | | | | meetings captured in videos, creating textual | | | | records of construction progress, challenges, | | | | and decisions.

a6070ed8-da91-4376-a551-f268427f1920

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.7. Video To Text | | records of construction progress, challenges, | | | | and decisions. | | | | Post-construction | Commissioning reports | Generate performance reports by transcribing | | functional testing/acceptance videos. | | | | All | Project Archive | Create searchable records of the whole | | project by transcribing videos into indexed | |

78d33019-c484-475f-ae62-c6c2dbd2205f

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.7. Video To Text | Project Archive | Create searchable records of the whole | | project by transcribing videos into indexed | | | | documentation. | | |

19cfc2fb-8102-4e72-8e4e-1c6b46c8398f

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.8. Video To Image As shown in **Table 19**, extracting key representative images from construction videos offers value across construction projects. Exploring site videos could be condensed into salient snapshots during pre-construction to accelerate assessments. Video conferences discussing design concepts can be automatically packed into a collage of snapshot visuals [108]. In the construction phase, delivery footage could be processed into consolidated photo logs of materials arriving on-site. Aerial construction video can generate periodic bird's-eye progress views [106]. Post-construction applications include extracting instructional stills from facility tutorials. Compiling time-lapse visual collages from archival videos can summarize entire project journeys. Video summarization techniques such as recurrent auto-encoders demonstrate capabilities to identify important frames that distill key visual concepts from longer video sequences [109]. Through the application of these techniques to construction footage, essential moments can be extracted without the necessity for manual video scrubbing. The representative thumbnail images could support rapid video review and summarization. They also integrate more seamlessly into reports and presentations compared to video embeds. Further innovation in dense video understanding and summarization will continue expanding the capabilities for automating the extraction of impactful visuals from construction multimedia. | Potential opportunity | Description | Project phase | |------------------------------------------------|-------------------------------|-----------------------------------------| | Pre-construction | Site exploration frame | | | extraction | |

28dd508b-ceb3-4e78-bd75-3ff249e6d395

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.8. Video To Image | | extraction | | | | Extract key frames from site exploration | | | | videos, creating static images that capture | | | | crucial moments and details for initial site | | | | assessments. | |

f7dbe55f-9f5a-4c40-a34b-b1efc1bfc517

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.8. Video To Image | | assessments. | | | | Pre-construction | Conceptual design snapshot | | | generation | | | | Extract representative snapshots from videos | | | | discussing conceptual designs and creating | | | | visual representations of architectural |

a0a9daf3-c780-4c4c-86d6-9a45e2e8ed2f

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.8. Video To Image | | discussing conceptual designs and creating | | | | visual representations of architectural | | | | concepts for documentation and | | | | presentations. | | | | Pre-construction | Virtual landscape preview | | | stills

9a8f593d-0c03-45cb-872e-485303d27e5a

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.8. Video To Image | | Pre-construction | Virtual landscape preview | | | stills | | | | Generate still images from videos showcasing | | | | virtual landscape previews, providing | | | | stakeholders with static visual references for | | | | pre-construction landscape assessments. |

4c4c7962-5d6d-4345-98d3-5ac38ad82bb4

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.8. Video To Image | | stakeholders with static visual references for | | | | pre-construction landscape assessments. | | | | Construction | Material delivery visual logs | Generate representative photo logs from | | videos capturing materials and equipment as | | | | they arrive on site. | | | | Construction | Remote and automated | | | progress

21636155-7e40-419f-9b15-7759aa789312

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.8. Video To Image | | Construction | Remote and automated | | | progress monitoring | | | | Aerial video can be processed to | | | | automatically produce interval imagery | | | | depicting bird's-eye views of the site at | | | | different points in time.

f404df29-428d-4aa5-95e7-325957476666

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.8. Video To Image | | depicting bird's-eye views of the site at | | | | different points in time. | | | | Post-construction | Facility usage instructions | | | still | | | | Extract still images from videos providing | | | | facility usage instructions, creating visual

8a7363a4-5967-4280-9e2b-83fbd4789a18

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.8. Video To Image | | Extract still images from videos providing | | | | facility usage instructions, creating visual | | | | stills that convey important guidelines for | | | | post-construction occupants. | | | | All | Comprehensive project | | | timeline collage

d0299353-92fa-4730-95ac-a6395ae5578b

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.8. Video To Image | | All | Comprehensive project | | | timeline collage | | | | Compile critical frames from videos across | | | | all project phases into a comprehensive | | | | timeline collage, visually summarizing the | | | | entire project journey. |

7d1ca50b-50ff-477e-b3dd-108d9bb9868f

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.8. Video To Image | timeline collage, visually summarizing the | | | | entire project journey. | | |

7319401f-407d-4179-8909-f6dcbc201da3

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.9. Video To Video While static outputs enable analysis, video can engage stakeholders through dynamic visualization. Constructing the future requires envisioning it in motion. Generative video-tovideo models can help bring these visions to life across the project lifecycle, as depicted in Table 20. During pre-construction planning and bidding, generative models could synthesize simulated construction sequences from source videos to allow interactive visualization of various work strategies and schedules for optimization [110]. On construction sites, input training videos could be adapted into multi-lingual versions translated across diverse crews to increase accessibility and comprehension. Architectural visualization videos, as-built documentation, and sensor data could be synthesized into lifecycle simulations and digital twin representations to support operations and maintenance. Opportunities exist across all phases, such as video quality enhancement, accelerated time-lapses, and video summarization for efficient review. By learning spatial-temporal relationships from construction footage, models can extend source videos into modified outputs adapted for downstream requirements. This allows tailoring visual media for specific applications ranging from training to monitoring to forecasting. As video generation techniques continue advancing, the potential for AI-assisted video remixing and synthesis to enhance multimedia value across the construction project lifecycle will grow substantially [97]. Processing datasets accumulating from the proliferation of construction cameras and sensors using generative video models promises to unlock new visual insights and perspectives. | Potential opportunity | Description | Project phase | |------------------------------------------------|---------------------------|-----------------------------------------------| | Pre-construction | Cost Estimation | Synthesize timelapse estimates of different | | construction methods/schedules from sourced | |

7ed27eff-5557-42a7-841f-caf20e705621

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.9. Video To Video | Cost Estimation | Synthesize timelapse estimates of different | | construction methods/schedules from sourced | | | | videos. | | | | Pre-construction | User Experience Preview | By generating composite videos from design | | concept footage, animations, and 3D models, | | | | the expected user experiences and functional | | | | flows within a proposed development can be |

99ecd11c-28cc-4990-89cb-12afb867129b

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.9. Video To Video | | the expected user experiences and functional | | | | flows within a proposed development can be | | | | simulated before construction. | | | | Construction | Safety Orientations | Simulate hazard scenarios for training by | | compositing archived incident videos. | | | | Construction | Multilingual Translation | Generate multilingual versions of | | instructional videos to support diverse crews. |

1ac90592-9df8-4fe9-9c15-de0e44642a06

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.9. Video To Video | | Construction | Multilingual Translation | Generate multilingual versions of | | instructional videos to support diverse crews. | | | | Post-construction | On-Boarding Video | Produce guided video facility tours from | | archival documentation for onboarding and | | | | handoff. | | | | Post-construction | Lifecycle Forecasting | Taking as-built

b664bf8c-c825-4ead-8ac3-ab6eab2799e1

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.9. Video To Video | | | | Post-construction | Lifecycle Forecasting | Taking as-built construction videos detailing | | "as constructed" conditions, future | | | | renovations, retrofits, refurbishments, or | | | | capital upgrades planned at different phases | | | | of the asset lifespan can be digitally | | | | prototyped and overlaid.

3ff89d63-9fbf-48e3-83b6-bc1992550ad0

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.9. Video To Video | | of the asset lifespan can be digitally | | | | prototyped and overlaid. | | | | All | Video Quality Enhancement | The upscale resolution, framerate, colors, | | etc., for legacy or damaged construction | | | | footage. | | | | All

8da82f05-248a-4056-b8aa-63ecc47a1d7a

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.9. Video To Video | | | | All | Accelerated Playback | Generate compressed timelapse videos from | | lengthy footage for rapid review. | | |

20723465-d417-4382-96c3-fee096975ea4

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.3. Challenges The adoption of generative AI in the construction industry is growing; thus, it comes with challenges. As presented by the experts during the Delphi survey, these challenges are multifaceted, encompassing domain-specific, technological, adoption, and ethical challenges, as shown in Figure 8. These challenges should be navigated for successful real-world deployment of generative AI in construction. A holistic approach considering technical and non-technical factors is required to overcome barriers and unlock the full potential of LLMs in construction.

f5f97b0f-e0fc-4fe8-8130-ad709e5ddd1e

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.3.1. Domain-Specific Challenges i. Requirement for construction-specific knowledge Construction is a complex field with intricate technical knowledge required to execute projects safely and efficiently [111]. However, most current generative AI techniques rely solely on statistical patterns extracted from data. They need the ability to explicitly encode the nuanced human expertise and domain constraints around structural engineering, materials science, construction codes, aesthetics, machinery, schedules, costs, etc. As a result, generative models trained exclusively on construction data may fail to produce valid, high-quality outputs that align with industry best practices. For example, a generative design model may create a visually appealing 3D building model that violates important structural principles, safety factors, or zoning regulations. The lack of engineering heuristics and constraints leads the unrestrained model to hallucinate flawed plans. Likewise, a generative text model trained only on construction documents will fail to generate specifications or instructions demonstrating a human's comprehension of materials compatibilities, sequences of operations, cost impacts, or equipment capabilities. Generative models need better integration of structured domain knowledge beyond just data patterns to reach their potential in construction. This remains challenging as industry experts' rules and mental models are difficult to codify for machines. Advances in neuro-symbolic AI, modular architectures, and expert-in-the-loop training show promise for imbuing models with more robust construction domain intelligence [112].

2b669b9b-7a49-432d-8b5e-7b1cfb9a7d02

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Ii. Handling Unstructured And Heterogeneous Data Construction data exists in multifaceted, unorganized formats across disparate systems, posing challenges for generative AI. Project information encompasses everything from scanned paper blueprints to 3D BIM models, permits, contracts, change orders, requests for information (RFIs), submittals, specifications, budgets, meeting minutes, multimedia, and more. These data types have different structures, semantics, units, symbols, file formats, and modalities. Generative models like GANs and VAEs struggle to ingest this heterogeneous, unstructured data directly to synthesize coherent outputs. For example, a basic image-to-image model cannot map a 3D BIM model, change order form, and permit application into a unified generated output. The variability across projects also hampers standardized tooling. Each construction firm may have customized data conventions, nomenclatures, templates, and systems tailored to their needs. Creating consolidated datasets from dispersed historical records is arduous. To work around these challenges, purpose-built multi-modal generative architectures are necessary [113]. Techniques like attention mechanisms, graph networks, and transformer models show promise for learning alignments and correlations across varied data types. However, no universal solution exists to handle the messiness of real-world construction data. Generative AI still requires extensive wrangling of unstructured inputs into tidy, normalized features.

f5b8945b-8ac2-471f-b366-4b37a07b0d51

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Iii. Lack Of Large-Curated Datasets Generative AI models need massive, high-quality training datasets to perform well. However, most construction firms do not consistently organize and consolidate their project data into formats usable for training a large model [103]. Historical records remain fragmented across various databases, file shares, and systems. The effort required to aggregate and clean unstructured construction data into coherent datasets is prohibitive without dedicated workflows. Data may reside in legacy formats. Important contextual links between related data points may be lost. The need for more data versioning, consistency, and curation poses challenges. For niche construction applications like generating site layouts or drywall specifications, virtually no large canonical datasets exist publicly to train models [114]. Collecting sufficient data from scratch requires substantial industry participation across firms. Annotation and labeling also necessitate scarce expert time. Without sizable, high-coverage training datasets, generative models struggle to generalize. They easily neglect sparse edge cases or unique scenarios found in complex construction projects. Models trained on inadequate data produce lower-fidelity outputs that lack realism and conformity to standards. Overcoming this bottleneck will require construction firms to systematically organize data accumulation, annotation, versioning, and consolidation workflows. Precompetitive industry data consortiums can also help aggregate datasets for typical AI applications.

78204f3c-68a0-4a1b-a871-36a50147415a

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Iv. Bias In Existing Datasets Construction datasets often exhibit significant regional biases propagated through generative AI models trained on this data. For example, architectural plans and building methods reflect local materials availability, weather patterns, seismic requirements, and zoning laws. Specifications follow jurisdiction-specific codes and standards. Units of measure, terminology, and language also vary geographically [115]. If models are trained solely on historical datasets from a particular country or city, the generated outputs will inherit these narrow perspectives. A design model trained only on American examples may overlook important considerations for cyclone-prone regions when deployed in the tropics. Likewise, language models trained only on specifications for a particular state could generate confusing RFI responses for international contractors following different norms. Outputs may also inadvertently include inapplicable regional colloquialisms. Training datasets must include wide diversity along multiple geographic axes to minimize bias to improve model coverage. However, thoughtfully collecting and curating such datasets is challenging for firms focused on their local region. Synthetic data augmentation techniques can help artificially expand variety once baseline data is available [116]. In practice, biased training sets often necessitate maintaining individualized models tailored to each application region. But this multiplicity hampers scaling and adds overhead. Developing adaptable generative models that generalize across diverse contexts remains an impactful challenge in construction.

0a87bed0-5c53-45e7-ab11-3b9ff0f10a37

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## V. Integration With Workflows And Standards The construction industry has relied heavily on workflows and proprietary systems tailored to each firm's needs and project requirements for many years. Seamlessly integrating generative AI solutions with these incumbent environments poses significant adoption difficulties. A core challenge is the need for interoperability between the modern ML tools underpinning generative models and the fragmented legacy software prevalent in construction. Custom integrations are needed to connect predictive models with databases, analytics dashboards, enterprise resource planning (ERP) platforms, BIM tools, and more [103]. However, construction systems often lack application programming interfaces (APIs). Generative models also need flexibility to adapt outputs to the proprietary data structures, nomenclatures, and templates used within each company. One-size-fits-all solutions struggle without customization. Firms are also reluctant to overhaul proven workflows solely to accommodate AI systems that appear disconnected from daily tasks. For adoption, generative models should directly build on available in-house data while aligning outputs to industry-standard specifications, equipment libraries, materials databases, regulations, and best practices. Workers are more inclined to use AI content that meshes with familiar domain paradigms rather than introducing foreign concepts. Overcoming these integration hurdles requires either extensive custom development efforts or architectures adaptable enough to map generative outputs to diverse construction environments out of the box. Finding the right balance between generalization and specificity remains an obstacle to embedding AI within incumbent workflows.

86047a27-3d5a-492b-8969-0fce7f7c1a34

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.3.2. Technological Challenges I. Model Instability And Training Difficulties Training generative AI models like GANs, GPTs, and LLMs to reliably produce stable, highquality outputs remains challenging, even in broader application areas outside construction. [117]. These training and stability issues become even more pronounced in the complex, constrained construction industry. The non-linear neural network architectures underlying many generative models have billions of parameters optimized through stochastic gradient descent [33]. The internal representations and dynamics of these massive models still need to be better understood, making their unpredictability harder to troubleshoot. During training, generative models are prone to problems like mode collapse, failing to capture the full diversity of training data. Finding the right balance between overfitting the data while still being able to generalize is tricky. Other issues, like vanishing gradients, can prevent networks from adequately learning. These training instabilities are amplified when models are scaled to handle sizeable multi-modal construction datasets. Getting models to synthesize completely novel outputs unrestrained by training patterns, as required in generative tasks, also increases unpredictability. Advances in principled network design, normalization techniques, robust optimization algorithms, and better training diagnostics should improve model stability. But for now, the opacity and fragility of uncontrolled generative synthesis pose an inherent challenge.

f1d3fda1-f0c6-4303-9f9d-2b1d1c0e0c8c

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Ii. Computational Resource Requirements Generative AI models are extremely computationally intensive, both during training and inference. State-of-the-art models like GPT-4 contain billions of parameters, requiring extensive parallel processing power on specialized hardware like GPU clusters or TPUs to train within reasonable timeframes [103]. For smaller construction firms, procuring and operating this expensive infrastructure may be infeasible just for experimenting with generative AI. Outsourcing to cloud platforms can mitigate costs but still demands significant investment. The carbon emissions footprint from model training should also be considered, given sustainability goals in the industry [118]. Even after models are trained, deploying them for inference and generating new outputs is resource-intensive. Real-time generation of high-resolution images, 3D models, or lengthy text would require low-latency access to powerful cloud computing. Many construction companies need more modern on-demand computing resources. As model sizes and demand for higher-quality outputs increase, so will hardware requirements. Construction firms without the IT infrastructure or budgets to continuously upgrade generative AI capabilities risk being left behind. This could create a bifurcation where only the most prominent players can afford to operate at the state-of-the-art. Advancement of more efficient architecture, distillation techniques, and on-device inference chips may eventually dampen costs. But in the interim, the level of resources needed to benefit from generative AI poses barriers, especially for smaller general contractors and subcontractors. Strategic partnerships with tech providers could help navigate the substantial computing investments involved. iii. Assessing output quality Unlike discriminative ML models, where accuracy metrics quantify performance, evaluating generative AI outputs' true quality is difficult. Metrics like Fréchet Inception Distance provide a proxy for similarity to accurate data distributions. However, these have limited utility when outputs are meant to be completely novel syntheses tapping the unknown. For niche construction applications, benchmarking datasets to test against do not exist. Assessing quality often relies on slow and subjective human review by domain experts, which does not scale. Furthermore, generated outputs like text, images, or 3D models may appear convincing on the surface to non-experts, exhibiting clear style and coherence [119]. However, upon closer expert inspection, these outputs lack deep domain-specific fidelity and violate constraints that may be obvious to a construction professional. Detecting these subtle faults, which do not manifest in surface metrics, remains an open problem. Developing and integrating better quality assurance techniques for generative AI in construction is

7e9ecb57-fd38-48eb-bf04-4b52b1c7be86

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Ii. Computational Resource Requirements construction applications, benchmarking datasets to test against do not exist. Assessing quality often relies on slow and subjective human review by domain experts, which does not scale. Furthermore, generated outputs like text, images, or 3D models may appear convincing on the surface to non-experts, exhibiting clear style and coherence [119]. However, upon closer expert inspection, these outputs lack deep domain-specific fidelity and violate constraints that may be obvious to a construction professional. Detecting these subtle faults, which do not manifest in surface metrics, remains an open problem. Developing and integrating better quality assurance techniques for generative AI in construction is crucial. This will likely require a combination of automated quantitative checks, qualification processes, and skilled human reviewers. Without rigorous validation protocols, using generative models for safety and cost-critical construction tasks is precarious. All stakeholders need reliable indicators that system outputs meet domain requirements before fully embracing generative techniques.

af1a3cdf-a7f9-4b31-8a16-1e7d35b82a28

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Iv. Potential For Hallucination And Factual Inconsistencies A significant danger posed by generative AI is its tendency to fabricate imaginary details that appear valid but diverge from reality[30]. When synthesizing novel outputs, these models are unrestrained by the fixed training data distribution. The systems "hallucinate" new content by stochastically combining learned features and patterns. In open domains like art and entertainment, such an unconstrained generation of new ideas may be desired. But for the safety-critical construction industry, factual inconsistencies or false details could have disastrous consequences if relied upon. Even minute defects in a generated building design, equipment specification, or work procedure could lead to accidents, delays, or rework down the line. Unlike discriminative models, which stick tightly to input features, generative models have free rein to distort outputs during synthesis. While coherence and surface plausibility remain high, factual correctness often suffers [103]. Without proper oversight, these distortions go unnoticed until problems arise in construction or operations. The unpredictable, unsupervised nature of generative AI makes it fundamentally risky for domains requiring tight conformance like construction. Extensive validation processes led by human experts and automated safety checks are necessary when applying these models. However, detecting the subtle faults unique to generative approaches remains an open research problem. Only when more controlled techniques are developed, unleashing unconstrained generative models comes with high uncertainty. Their propensity to smoothly fabricate imaginary details outside the training distribution should instill caution. While promising, balancing generative AI's creative potential with construction constraints is critical.

3901f792-7bcb-4f8e-981c-9ad0cb63292e

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## V. Lack Of Explainability A significant limitation of modern generative AI techniques is their black-box nature. Systems like LLMs and GANs offer minimal transparency into their internal reasoning for producing specific outputs over others [120]. The models synthesize outputs by propagating input signals through billions of transformations across neural network layers. Explaining why one output manifested versus another is nearly impossible given this complexity. In construction, lack of explainability poses risks and makes diagnosing errors harder. When designs, images, or text are generated, professionals have no visibility into the generative model's intent or rationale. This needs to be revised in order to maintain human oversight of the system's thinking and conclusions. If flaws are detected, the opaque models provide little clue into the root causes. Troubleshooting and correcting errors becomes a guessing game without explanatory abilities. This could lead to blind trial-and-error tuning versus informed debugging. More transparent and controllable architectures may be needed for broader acceptance in the relatively conservative construction industry. Hybrid approaches combining neural networks with declarative knowledge about engineering constraints could improve interpretability. Interactive interfaces that allow step-by-step manipulation of generative models also offer more transparency.

de673af3-0ef2-4cc3-bcfc-c2fe39eced32

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.3.3 Adoption Challenges I. Resistance To New Technologies The construction industry has historically needed to be faster to adopt new technologies compared to other sectors. This inertia and resistance to change stems from several interrelated factors. Many construction firms rely heavily on established processes and workflows that have been incrementally optimized over the decades. There is often a reluctance to modify or replace these proven legacy, deeply ingrained methods [121]. Furthermore, the supply chain involves disparate stakeholders with different capabilities and resources. Aligning on new technology adoption is difficult across this fragmented ecosystem. At a management level, there are concerns that AI could disrupt traditional roles and ways of doing business in construction. The industry relies on specialized trades and processes that workers have invested years into mastering. Introducing unfamiliar systems feels inherently risky, making management hesitant to champion large-scale technology overhauls. Overcoming these barriers will require a combination of peer-based advocacy, demonstratable benefits, incentives, change management planning, and strong leadership buy-in.

c53ac8f2-31a6-4c8e-8ae1-2a0ed02a6b53

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Ii. Lack Of Skills And Expertise The use of generative AI requires specialized skills that currently need to be improved in most construction companies [30]. While these firms have deep domain expertise in construction processes, materials, equipment, etc., they have limited in-house experience with AI and data science. Most construction companies cannot realistically build large internal AI teams from scratch. Construction firms will likely need to hire dedicated AI talent or partner with technology firms to complement their domain knowledge. However, professionals with deep AI expertise and construction industry knowledge are rare and difficult to recruit. Closing the skills gap will require a combination of recruitment, training, partnerships, and creating more no-code or low-code solutions tailored to the industry.

825c05be-caeb-4b09-a61e-367c7bd615a6

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Iii. High Upfront Investment Costs Adopting generative AI poses considerable upfront investment costs, which may deter construction firms from pursuing it. Firstly, data preparation requires aggregating dispersed historical data from multiple systems and getting it into a unified format [122]. Next, licensing and developing generative models necessitates paying for specialized AI services. The computational resources needed for training and inference, such as cloud GPUs, add to the technology bill. Integrating the AI system with existing construction workflows and IT infrastructure demands custom development efforts [103]. Finally, machine learning engineers incur ongoing maintenance, monitoring, and enhancement costs. For large construction corporations, these expenses may be feasible to absorb. However, smaller contractors and trade firms operate on tighter margins and budgets. Many may find the capital expenditures required to implement generative AI prohibitively high. The construction industry needs to be more fast-moving initiatives on investments, especially for emerging technologies like generative AI. Demonstrating a convincing return on investment is critical for securing buy-in.

bebe9dbf-a944-473f-831f-5a4d4e12f54b

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Iv. Immature Supporting Infrastructure Successfully implementing generative AI requires data infrastructure and workflows, which are currently immature in the construction industry. Firstly, most firms lack the data pipelines and consolidation needed to feed massive training datasets to generative models. Data labeling and annotation workflows necessary for supervision are also non-existent. Furthermore, the machine learning operations (MLOps) tools for versioning models, monitoring systems, and ongoing improvement are foreign to most construction IT departments. Generative AI relies extensively on the scale of computing power, demanding integration of construction data systems with cloud platforms [14]. However, seamless connections between internal databases, BIM models, and external cloud resources are rare [123]. There is also a shortage of prebuilt integrations between construction software tools and generative AI APIs. The surrounding ecosystem to enable enterprise adoption is still evolving. In effect, construction firms cannot simply plug and play off-the-shelf generative AI solutions into their existing IT systems. Substantial infrastructure development and integration efforts are required to create the data and compute foundations. For many companies, this necessitates a complete overhaul of internal data practices, development stacks, and system architectures.

392760fa-19f7-4965-8ebf-188501ad92b3

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## V. Unclear Governance Frameworks There are unresolved questions around legal liability - who is accountable if an AI system produces faulty designs, specifications, or recommendations that lead to accidents? Quality control and validation protocols for generative models in construction are also lacking. Furthermore, the security implications of relying on AI to guide mission-critical construction processes are still being worked out. Risk management frameworks and technical standards have not caught up to the rapid advances in generative techniques. There are also ethical concerns about reproducing historical biases in data, which require governance to be addressed transparently and responsibly. The regulatory regime surrounding generative AI in construction is unclear and fragmented. Companies are hesitant to deploy unproven technologies without best practices or precedents to follow. Both public and private institutions need clear legal guidelines, technical validation protocols, model risk management expectations, and standards of use [124].

30f6006d-1774-4b44-9851-ab60ef7f06a7

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.3.4 Ethical Challenges I. Data Privacy And Security Construction projects generate vast amounts of potentially sensitive data - from financial records to design specifications to site photographs. As this data is increasingly used to train generative AI models, firms must act responsibly to respect privacy and maintain trust [122]. However, most construction data practices are focused on operations rather than ML readiness. Efforts will be needed to obtain proper consent, audit datasets, and implement access controls for AI systems. Data anonymization techniques can help remove personally identifiable information. But details like project names, locations, and dates often cannot be fully stripped without losing utility. Strict governance models for internal data collection, external usage, and retention will need to be developed. Cybersecurity is also critical, given the highly sensitive nature of commercial construction data. Breaches during model development or deployment could have serious consequences ranging from confidentiality violations to industrial espionage. Construction firms can uphold privacy while tapping AI advancements by minimizing risks through responsible data curation, anonymization where possible, and tight access restrictions. However, this may require overhauling ingrained data practices focused on operational efficiency and ethics. The cultural and procedural shifts will challenge organizations to harmonize AI progress with core principles of trust and transparency [125].

04ce1755-3b97-4e54-b81c-2d8a5339e39a

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Ii. Social Concerns About Job Automation With the potential to enhance many human tasks, the adoption of generative AI in construction raises understandable concerns about workforce impacts. However, the effects are unlikely to be straightforward substitution of workers. AI may automate narrow, repetitive tasks but augment professionals to be more productive on complex strategic initiatives [126]. New human roles overseeing and collaborating with AI systems will also emerge. Proactive communication, training programs, and organizational change management will be imperative for a responsible transition. Leaders must be cognizant of apprehensions among workers fearing replacement by "thinking machines." Construction firms that are reliant on specialized trades have a particular responsibility to involve and support affected staff through an AI- enabled transformation. Instead of blunt displacement, AI should aim for symbiosis - enhancing professionals' capabilities while handling rote work. Adoption with the right intention of uplifting workers and augmenting expertise can help construction firms achieve societal benefits and sustainable competitiveness.

25382210-dc3f-4cb4-ae49-05eb2c0c3f95

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Iii. Potential For Misuse The autonomous and scalable capabilities of generative models create risks of misuse if deployed irresponsibly. For instance, AI systems lacking appropriate safeguards could generate realistic but structurally flawed building or equipment designs. Without rigorous engineering constraints and oversight, the unrestrained creativity of generative models could produce designs that circumvent safety codes and regulations [119]. Similarly, project plans, budgets, certificates, invoices, change orders, and other documentation falsified by AI could enable fraud or errors. The ease of generating convincing paperwork at scale for malicious purposes poses financial and legal risks. To prevent misuse, construction firms need to implement extensive technical and ethical precautions [127]. This includes carefully auditing training data and models for issues like bias, establishing sandboxed development environments, verifying outputs, and instituting human-in-the-loop checks before deployment. Responsible governance encompassing explainability, transparency, and accountability is also critical. Generative AI offers immense opportunities but also risks if its capabilities are unleashed carelessly. With prudent controls and oversight, construction professionals can minimize hazards while benefiting from accelerated innovation.

f35de33c-0408-4d37-8572-3bdaafa3f168

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 5. Framework For Building Custom Lgm In The Construction Industry While general pre-trained models like GPT-4 and Gemini Pro offer promising capabilities [45,95], developing LGMs customized for the construction domain can further enhance performance on industry-specific tasks. This section provides a framework that construction professionals and firms can follow to develop tailored LGMs using their proprietary data (see Figure 3). The key steps include construction data collection, dataset curation, training of custom LGM, evaluation of the LGM, and deployment.

9269f857-52e9-4789-9856-ca8aa303bc93

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 5.1. Construction Data Collection The first critical step in developing a custom LGM for construction is aggregating a broad corpus of relevant data from past projects and documentation across the firm. This serves as the foundation for training the model to comprehend and generate high-quality industryspecific language [103]. The data should be pulled from a diverse range of historical sources to cover the full breadth of concepts and terminology used within the company's work. Potential sources that should be tapped include technical specifications, equipment manuals, permit applications, contractor invoices, design reports, construction schedules, requests for information, project budgets, safety protocols, inspection checklists, as-built drawings and videos, relevant codes and standards, project contracts, and meeting minutes [86]. Essentially, all unstructured data around projects both directly produced by the firm and exchanged with partners, contains valuable language samples that can educate the LGM. Ideally, the data collection should draw from both successful and problematic construction projects within the company's archives. This provides balanced examples and helps the model better handle edge cases by learning from challenging historical incidents. Maximum diversity in the kinds of projects covered also allows the LGM to generalize robustly. In terms of format, the text data should be structured into machine-readable JSON if readily available in this format within the firm's document systems [26]. However, extensive cleaning and preprocessing of diverse unstructured data will likely be required. For scanned or image-based data, optical character recognition can extract text. For legacy video and audio, speech recognition techniques can generate transcripts. Metadata extraction can pull useful tags and descriptions from media files. Point cloud data may need processing into voxel grids or meshes. The raw data extracted across modalities like text, image, video, and audio must then be transformed into standardized corpora in formats digestible for model training. Expect substantial effort to munge multifaceted data sources into shapes consumable by generative algorithms.

dd2548da-a84e-41be-a15e-90dcbb8a5ebd

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 5.2. Dataset Curation After aggregating raw data, the next step is carefully curating it into a high-quality dataset ready for training the construction LGM. This involves extensive processing and analysis. First, any sensitive personal information or proprietary business data must be removed from the corpus to respect privacy and security protocols. Next, identifiable entities like specific project names and locations should be anonymized where possible to mitigate risks. Thorough cleaning is required to fix any formatting inconsistencies, OCR errors, or annotation issues so that the data is pristine. The data sources should also be analyzed to ensure sufficient diversity - if the dataset focuses too heavily on certain project types or documentation formats, it can lead to a lack of broad applicability. Chronological splitting into train, validation, and test sets is also critical for properly evaluating model performance over time [15,32]. Moreover, synthesizing additional diverse examples through techniques like contextual data augmentation should be considered to boost the coverage of niche cases. Domain experts should manually sanity-check random samples from across the final dataset to catch any lingering issues before training begins. This human-in-the-loop auditing step provides quality control and ensures the data is aligned with true construction language [128].

eb2cb8db-cae8-487b-8c33-2f1813059c49

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 5.3. Training Custom Lgm With a tailored construction-specific dataset prepared through careful data curation, the next phase is leveraging this data to train a custom LGM for construction tasks. The model can be initialized from scratch and trained end-to-end on just the domain data. However, it is also common to initialize with an existing general pre-trained LGM like GPT-4 that has already learned strong linguistic representations from web-scale corpora [45]. This foundation can then be fine-tuned on the construction dataset to adapt to industry-specific terminology and patterns. Transfer learning in this manner can significantly reduce the computational resources and time required for training versus a from-scratch approach [87]. Regardless of the initialization technique, the overall training methodology involves first selecting an appropriate underlying model architecture and size. Transformer networks currently demonstrate state-of-the-art performance on language tasks but require tuning of their complex configurations to fit each dataset and use case [46]. Training of the model is then conducted using GPU or TPU computational infrastructure until convergence on the construction training data distribution as measured by validation set performance. The training hyperparameters, including batch size, learning rate schedules, and activation functions, must be finely tuned based on the validation results to optimize model quality. This entire model development process requires extensive computational resources for both the core training and the surrounding hyperparameter optimization. An opportunity to enhance the training using retrieval augmented generation (RAG) also exists [129]. A retrieval system can be built on top of the construction dataset to provide contextual examples that keep the LGM grounded in domain-specific language during training. Hence, developing a high-quality, performant LGM customized to construction data involves synthesizing diverse techniques into a robust training methodology.

2b4f3e9b-fa4e-4b97-9ebb-db95b9b00b1d

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 5.4. Lgm Evaluation Once the LLM is trained on construction data, it must undergo rigorous evaluation before deployment to ensure it achieves the performance and quality thresholds required in downstream applications. The semantic coherence, grammar, terminology, and validity of generated outputs should be extensively assessed via qualitative human review by domain experts. This allows for validating that the model produces high-quality language aligned with true construction concepts. Checking for potential biases and factual inaccuracies is also critical to avoid operational risks. In addition, the model should be quantitatively benchmarked against baseline methods on domain-specific tasks using relevant metrics. For instance, the customized LGM can be evaluated for construction project phase classification accuracy and compared to off-the-shelf generic models. Other quantitative tests might include cross-referencing generated project budgets against actual data to assess fidelity [130]. Any shortcomings identified during evaluation should be addressed by re-training the model using modified data that improves coverage of deficient areas or adjusting the model architecture itself. The evaluation process also provides feedback for additional training to continue enhancing the LGM post-deployment.

22bf9582-e756-42c4-8f1d-1d8b594410d3

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 5.5. Custom Lgm Deployment To enable scalable deployment, the model should first be containerized using technologies like Docker and Kubernetes [131]. This encapsulates the model in a portable package with libraries and dependencies while managing computational resources. Exposing a high-performance API or web interface allows sending inference requests to the containerized model server. This powers integration into downstream domain applications. For instance, the custom LGM could be embedded within AI assistants, document generators, project recommendation systems, and other construction tools that benefit from its specialized text generation capabilities. Robust MLOps processes need to be implemented for continuous monitoring, versioning, and improvement of the model post-deployment [132]. As new project data comes in, it can be used to further tune and enhance the model to stay up to date. Human oversight and governance are critical during deployment to ensure quality control and responsible privacy, security, and ethics practices. With the proper infrastructure and processes in place, the custom construction LGM can be sustainably integrated to augment a wide range of business functions with an industry-tailored AI generation solution. This framework provides a methodical blueprint for construction firms to transform their data into strategic generative capabilities.

cdd19243-06f6-4c5f-924e-ac6ab151201c

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 6. Case Study - Information Retrieval And Knowledge Discovery To validate the potential for using generative AI in construction, a case study was conducted focused on information retrieval and knowledge discovery. This is one of the potential opportunities identified in the previous section on text-to-text applications. Querying contract documents is a valuable application, as contract documents contain critical project requirements and details but can be lengthy and complex to manually search through. The contract document employed for this case study was obtained from a consultancy firm that served as the project manager. The project entailed the construction of a three-story hostel facility for a higher education institution. The contract document contained key information on project scope, specifications, materials, timelines, costs, quality standards, and other crucial parameters. As contract documents like this are often dense and unstructured, retrieving information requires tedious manual searching. Generative AI offers the ability to query the document in natural language and receive direct answers summarizing the most relevant details. This case study demonstrates the value of training generative models on real-world construction contracts to improve information search and extraction.

84526ec8-660c-4683-ae43-79033a66e0f7

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 6.1. Model Development For this case study, the GPT-4 model was leveraged as the base LLM. GPT-4 is a proprietary LLM designed by OpenAI to generate coherent and useful text in a wide variety of domains [95]. It was pretrained on massive text datasets encompassing diverse topics and demonstrated state-of-the-art natural language processing capabilities [133]. To further enhance the capabilities of GPT-4 for the construction contract domain, a RAG framework was implemented on top of the base model. RAG integrates semantic search over a domain-specific knowledge base into the text generation pipeline. This allows retrieving the most relevant contextual examples from the contract text to prime the LLM when responding to queries. The RAG framework helps ground the model's outputs in the actual contract content, avoiding hallucinations. As depicted in Figure 10, the RAG pipeline consisted of [134]: - **Importing the contract document**: The first step was ingesting the raw text data from the contract document into the RAG system. This included preprocessed documents in PDF format. - **Splitting documents into coherent chunks**: The full contract document was segmented into smaller chunks of text spanning 3-5 sentences focused on a coherent part of the contract. This allowed more fine-grained contextual retrieval. - **Creating embeddings for the chunks**: ML embeddings were generated using an advanced semantic encoding model for each text chunk. OpenAI embedding was used for this purpose via API access. - **Storing the chunk embeddings in a vector knowledge base**: The chunk embeddings were indexed in a high-performance vector database. Cassandra database was used for this purpose [135]. This enables quick retrieval of contextually similar chunks. - **Accepting user query as input**: At inference time, the user provides a text query expressing their desired contract document information need. - **Embedding query into same vector space**: The input query is encoded into the same semantic vector space as the chunks using the same sentence model. - **Performing semantic search to identify relevant specification chunks**: Efficient approximate nearest neighbor (ANN) search is run to find chunks with the highest semantic similarity to the query vector. - **Ranking retrieved chunks by semantic similarity**: The topmost similar chunks are ranked and filtered to create a subset most relevant to the query. - **Prov

97402c15-934e-4906-b997-e28af6af17f9

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 6.1. Model Development the user provides a text query expressing their desired contract document information need. - **Embedding query into same vector space**: The input query is encoded into the same semantic vector space as the chunks using the same sentence model. - **Performing semantic search to identify relevant specification chunks**: Efficient approximate nearest neighbor (ANN) search is run to find chunks with the highest semantic similarity to the query vector. - **Ranking retrieved chunks by semantic similarity**: The topmost similar chunks are ranked and filtered to create a subset most relevant to the query. - **Providing top k chunks to guide LLM text generation**: The top-ranking contract document chunks are provided as contextual examples to prime the LLM to generate focused and valid contract text.

c729807a-c052-4883-9b39-6ce6f12613e9

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 6.2. Model Evaluation To develop an appropriate set of queries for evaluating the model, 30 potential natural language questions were initially formulated based on information contained within the contract document. These draft questions were reviewed by a panel of three experts with professional construction knowledge to validate that they represent realistic queries of interest to industry practitioners accessing such a document. The first expert confirmed 26 of the proposed questions as relevant, while the other two experts each validated 23 questions. By taking the intersection, a final set of 20 common questions validated by all three experts was derived. This cross-validated question set encompasses diverse query types covering key information needs that construction professionals would seek to retrieve from contract documents. The 20 expertapproved questions were employed to evaluate the models' performance at extracting relevant answers from the contract. The approved questions and contract document were provided to 3 experts from the original panel that validated generative AI opportunities and challenges. These experts evaluated the model's responses to each question based on four metrics, which are similar to the metrics adopted by Wolfel et al. [136]: - Answer - Assesses if the model provides a substantive response ("Yes") or a nonanswer like "I don't know" ("No"). - Quality - Rates the truthfulness of the response on a 5-point scale. - Relevance - Rates how relevant the response is to the query and contract on a 5-point scale. - Reproducibility - Assesses the consistency of responses to the same question on a 5- point scale. The average score (mode was employed for the "answer" metric) for each question was calculated across raters for both the baseline GPT-4 model and the GPT-4 plus RAG system. The results are shown in Table 21, enabling a quantitative comparison of the two models' performance in extracting accurate, relevant information from the contract document when queried in natural language. According to Table 21, the high answer rate of 100% for GPT-4 indicates it consistently provides substantive responses to the questions rather than failing to generate any reply. However, the lower quality score of 3.87 reveals some responses fabrication details are not actually present in the contract, as the model hallucinates plausiblesounding but incorrect information. The decent relevance rating of 4.01 shows GPT-4's outputs are topically on point but strained by invented content. The reproducibility score of 4.53 suggests some inconsistency across repeated queries as well. In comparison, the RAG- enhanced GPT-4 model achieves higher

1589a815-5756-42f0-a6c2-188b65843fd0

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 6.2. Model Evaluation Table 21, the high answer rate of 100% for GPT-4 indicates it consistently provides substantive responses to the questions rather than failing to generate any reply. However, the lower quality score of 3.87 reveals some responses fabrication details are not actually present in the contract, as the model hallucinates plausiblesounding but incorrect information. The decent relevance rating of 4.01 shows GPT-4's outputs are topically on point but strained by invented content. The reproducibility score of 4.53 suggests some inconsistency across repeated queries as well. In comparison, the RAG- enhanced GPT-4 model achieves higher quality and relevance ratings of 4.13 and 4.48, demonstrating improved faithfulness through grounding outputs in retrieved contract passages. This reduces hallucinated content substantially. The superior reproducibility of 4.77 also highlights more stability from RAG's contextual retrieval. However, the lower 90% answer rate points to limitations in linking some questions to pertinent evidence, causing the model to default to "I don't know" non-answers. The quantitative metrics illustrate RAG's ability to enhance faithfulness and mitigate risks of hallucinations that generative models like GPT-4 exhibit. | Model | Answer Quality (1 | Question | |--------------|----------------------|-------------| | Number | -5) | | | 1 | GPT-4 | Yes | | 2 | Yes | 5.00 | | 3 | Yes | 1.00 | | 4 | Yes | 2.67 | | 5

dc946720-e560-45b1-bb38-32ffc9d23dd1

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 6.2. Model Evaluation Yes | 5.00 | | 3 | Yes | 1.00 | | 4 | Yes | 2.67 | | 5 | Yes | 1.33 | | 6 | Yes | 5.00 | | 7 | Yes | 5.00 | | 8 | Yes | 4.33 | | 9 | Yes | 5.00 | | 10 | Yes | 3.00 | | 11 | Yes | 5.00 | | 12 | Yes | 5.00 | | 13 | Yes | 4.

6246d6f7-4dd8-47b7-8a64-f6bcd322ea22

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 6.2. Model Evaluation | | 11 | Yes | 5.00 | | 12 | Yes | 5.00 | | 13 | Yes | 4.33 | | 14 | Yes | 5.00 | | 15 | Yes | 1.67 | | 16 | Yes | 4.33 | | 17 | Yes | 5.00 | | 18 | Yes | 4.00 | | 19 | Yes | 1.67 | | 20 | Yes | 4.00 | | Average | | | | (Percentage) |

1637c9fa-7580-49d5-9b52-7edeaa691f12

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 6.2. Model Evaluation | 1.67 | | 20 | Yes | 4.00 | | Average | | | | (Percentage) | | | | | 100% | 3.87 | | (77.4%) | | | | | | | | GPT 4 + RAG | | | | | | | | 1 | | Yes | | 2 | Yes | 5.00 | | 3 | No

c9b29c69-afe9-481b-a064-dfd412faac6d

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 6.2. Model Evaluation | | 1 | | Yes | | 2 | Yes | 5.00 | | 3 | No | - | | 4 | Yes | 3.67 | | 5 | Yes | 5.00 | | 6 | Yes | 5.00 | | 7 | Yes | 5.00 | | 8 | Yes | 5.00 | | 9 | Yes | 4.33 | | 10 | Yes | 2.33 | | 11 | Yes | 1.67 | | 12

b0a50d80-cbea-43ed-90bc-82461566581f

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 6.2. Model Evaluation | 4.33 | | 10 | Yes | 2.33 | | 11 | Yes | 1.67 | | 12 | Yes | 5.00 | | 13 | Yes | 4.67 | | 14 | Yes | 4.00 | | 15 | Yes | 4.00 | | 16 | Yes | 1.67 | | 17 | Yes | 4.00 | | 18 | Yes | 5.00 | | 19 | No | - | | 20 | Yes | 5.00

8fea88b4-fd5e-4bee-827a-49a795a5d3be

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 6.2. Model Evaluation | Yes | 5.00 | | 19 | No | - | | 20 | Yes | 5.00 | | Average | | | | (Percentage) | | | | | 90% | 4.13 | | (82.6%) | | | | | | | Relation (1 - 5) Reproducibility (1 -5) 4.01(80.2%) 4.53 (90.6%) 4.48 (89.6%) 4.77 (95.4%) Figure 11 provides example query-response screenshots for questions 6 and 15, comparing the baseline GPT-4 model and the GPT-4 plus RAG system. Figures 11a and 11d show the GPT- 4 responses to these questions, while Figures 11b and 11e display the responses augmented by the RAG retrieval pipeline. Figures 11c and 11f highlight the relevant passages containing the answers in the original contract document. Examination of

d0411f09-7b1f-4458-bdd4-1cf09675f20b

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 6.2. Model Evaluation ) 4.01(80.2%) 4.53 (90.6%) 4.48 (89.6%) 4.77 (95.4%) Figure 11 provides example query-response screenshots for questions 6 and 15, comparing the baseline GPT-4 model and the GPT-4 plus RAG system. Figures 11a and 11d show the GPT- 4 responses to these questions, while Figures 11b and 11e display the responses augmented by the RAG retrieval pipeline. Figures 11c and 11f highlight the relevant passages containing the answers in the original contract document. Examination of the examples illustrates that both GPT-4 and GPT-4 + RAG correctly answered question 6. However, for question 15, GPT-4 hallucinates details about GCC Clause 44 that are not contained in the actual contract, while GPT-4 + RAG grounds its response in the original document context to extract the price adjustment formula accurately. This side-by-side comparison of real queries demonstrates how RAG augmentation improves faithfulness by reducing hallucination and retrieving corroborating evidence to support generative outputs.

07db967b-6415-4ef0-92df-9741953c346c

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## (a) (b) (c) (d) (e) (f)

f46c2d58-4d1d-4ca0-bc44-06ee36db88ab

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 6.3. Model Limitation While the retrieval-augmented GPT-4 model shows promising results in querying the construction contract document, some limitations need to be acknowledged. The model struggled to retrieve relevant passages for two of the questions, defaulting to uninformative "I don't know" non-answers. This indicates that the chunking strategies and semantic search techniques used were unable to adequately link some complex questions to supporting evidence in the contract document. Future studies can explore different embedding models and vector databases and compare their performance. Generalization is another limitation - the model was trained on just a single contract document and may fail to transfer to new projects with different terminology, formats, and content. Training on a large corpus of diverse contracts would likely improve out-of-domain robustness.

1863b885-0d2e-4b4e-a377-dba20a4cc58d

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 7. Conclusion This research aimed to provide a comprehensive analysis of the current state, opportunities, and challenges of applying generative AI in the construction industry. Key insights were synthesized through a systematic literature review and expert Delphi study. The literature review revealed that generative AI adoption in construction is still in very early stages, with just a handful of initial studies exploring applications like information retrieval, project planning, hazard recognition, and risk assessment. However, the great potential of generative techniques like large language models was highlighted for enhancing productivity, accuracy, and automation across construction tasks. The expert panel discussions further expanded on promising applications of generative AI in the construction industry during the pre-construction, construction, and post-construction phases. Opportunities were identified for major data types, including text, images, and video. At least seven potential opportunities for each data type were identified. For instance, the identified opportunities for image-text application include land survey data extraction, 3D model specification, blueprint digitization, daily progress image analysis, material quality assessment, inventory management, as-built documentation text extraction, warranty claim documentation, and visual inspection reports. The experts also outlined critical challenges that need to be addressed regarding domain knowledge, data, training, validation, integration, adoption, resources, and responsible governance. A methodology was proposed to guide construction professionals in building customized generative AI solutions using their own proprietary data. The framework steps of data collection, curation, model development, evaluation, and deployment aim to make these powerful technologies more accessible for practical industry deployment. The value of the framework was demonstrated through a case study on applying generative models for enhanced querying of construction contract documents. The retrieval-augmented system (RAG) showed a significantly improved ability to extract accurate, relevant information from contracts through natural language queries compared to a baseline generative model (GPT-4). In terms of quality, relevance, and reproducibility, the RAG system outperforms the base GPT-4 model by 5.2, 9.4, and 4.8%, respectively. While this study provides valuable insights into the application of generative AI in construction, certain limitations present opportunities for future work. The systematic literature review was confined to three databases - Scopus, Web of Science, and ScienceDirect - based on their structured interfaces and indexing criteria. Despite snowball searching, some relevant articles may have been missed by focusing on these sources. The expert Delphi panel size was also restricted due to resource constraints, although care was taken to recruit highly

bc44935d-630e-44c9-b213-e3387b860951

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 7. Conclusion , and reproducibility, the RAG system outperforms the base GPT-4 model by 5.2, 9.4, and 4.8%, respectively. While this study provides valuable insights into the application of generative AI in construction, certain limitations present opportunities for future work. The systematic literature review was confined to three databases - Scopus, Web of Science, and ScienceDirect - based on their structured interfaces and indexing criteria. Despite snowball searching, some relevant articles may have been missed by focusing on these sources. The expert Delphi panel size was also restricted due to resource constraints, although care was taken to recruit highly experienced professionals. For the case study, only a single base large language model and embedding technique were utilized due to API access costs. Testing multiple state-of-the-art models could reveal further performance gains. Future studies can build on these findings by expanding the literature review across more databases, recruiting larger expert panels as generative AI becomes more prominent in the construction industry, and experimenting with diverse generative architectures and embedding methods given sufficient computing power and budgets. Addressing these limitations represents an avenue for additional confirmatory research and comparative assessment on applying generative AI for construction tasks. Nonetheless, this study provides a solid foundation of insights and a practical framework to guide further advancement.

4b9794f9-025d-4ec3-8ad7-d42bca2b0e04

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Acknowledgment This research is supported by the Department of Building and Real Estate, The Hong Kong Polytechnic University, Hung Hom, and the Centre for Advances in Reliability and Safety (CAiRS), Hong Kong Science Park12/F, Building 19W, Pak Shek Kok, NT, Hong Kong, China.

2406cce4-1e97-46cd-8b0e-6f5c101c77be

# Llasmol: Advancing Large Language Models For Chemistry With A Large-Scale, Comprehensive, High-Quality Instruction Tuning Dataset Botao Yu Frazier N. Baker * Ziqi Chen * Xia Ning Huan Sun The Ohio State University {yu.3737, baker.3239, chen.8484, ning.104, sun.397}@osu.edu https://osu-nlp-group.github.io/LLM4Chem/

464bcf6b-d4f0-4594-92d8-937f78270726

# Llasmol: Advancing Large Language Models For Chemistry With A Large-Scale, Comprehensive, High-Quality Instruction Tuning Dataset ## Abstract Chemistry plays a crucial role in many domains, such as drug discovery and material science. While large language models (LLMs) such as GPT-4 exhibit remarkable capabilities on natural language processing tasks, existing work shows their performance on chemistry tasks is discouragingly low. In this paper, however, we demonstrate that our developed LLMs can achieve very strong results on a comprehensive set of chemistry tasks, outperforming the most advanced GPT-4 across all the tasks by a substantial margin and approaching the SoTA task-specific models. The key to our success is a large-scale, comprehensive, high-quality dataset for instruction tuning named SMolInstruct. It contains 14 meticulously selected chemistry tasks and over three million high-quality samples, laying a solid foundation for training and evaluating LLMs for chemistry. Based on SMolInstruct, we fine-tune a set of open-source LLMs, among which, we find that Mistral serves as the best base model for chemistry tasks. We further conduct analysis on the impact of trainable parameters, providing insights for future research.1

614eb663-74fd-482e-933a-e0073e845610

# Llasmol: Advancing Large Language Models For Chemistry With A Large-Scale, Comprehensive, High-Quality Instruction Tuning Dataset ## 1. Introduction Chemistry is a fundamental science that underpins countless aspects of modern life, ranging from drug discovery and materials science to energy production. To facilitate research and applications in this domain, deep learning models including graph neural networks (Kipf & Welling, 2017) and Transformer-based models (Vaswani et al., 2017) have *Equal contribution 1Dataset and model will be released at https://github. been developed for various chemistry tasks such as forward reaction prediction, retrosynthesis, property prediction (Schwaller et al., 2019; Zhong et al., 2022; Chen et al., 2023; Zhou et al., 2023). However, these models are usually taskspecific models, which neglect shared chemistry knowledge across tasks and can hardly be adapted to different tasks. On the other hand, large language models (LLMs) such as GPT-4 (Achiam et al., 2023), Llama series (Touvron et al., 2023a;b), and Mistral (Jiang et al., 2023) have emerged as general-purpose foundation models and demonstrate remarkable abilities on various natural language processing tasks (Chang et al., 2024; Thirunavukarasu et al., 2023; Yue et al., 2023; Zhang et al., 2023; Deng et al., 2023). However, when applied to chemistry tasks, LLMs show only limited capabilities (Jablonka et al., 2022; Guo et al., 2023; Hatakeyama-Sato et al., 2023). For example, Guo et al. (2023) conducted evaluations on eight chemistry tasks and observed that while GPT-4 outperforms other closed- and open-source LLMs, its performance is not comparable with that of task-specific deep learning models. Particularly, they found that GPT models perform poorly when a precise understanding of SMILES (Weininger, 1988), a widely used textual representation for molecules, is required. In addition to directly applying pretrained LLMs, Fang et al. (2023) fine-tuned LLMs on an instruction tuning dataset, but their performance remains very low, far behind the state-of-theart (SoTA) models designed and trained for specific tasks. Given the discouraging results thus far, some critical questions arise: Are LLMs actually able to effectively perform chemistry tasks? Or, Are they fundamentally limited

941f4905-6d01-4e3a-ac4c-eb53a1bfbf12

# Llasmol: Advancing Large Language Models For Chemistry With A Large-Scale, Comprehensive, High-Quality Instruction Tuning Dataset ## 1. Introduction specific deep learning models. Particularly, they found that GPT models perform poorly when a precise understanding of SMILES (Weininger, 1988), a widely used textual representation for molecules, is required. In addition to directly applying pretrained LLMs, Fang et al. (2023) fine-tuned LLMs on an instruction tuning dataset, but their performance remains very low, far behind the state-of-theart (SoTA) models designed and trained for specific tasks. Given the discouraging results thus far, some critical questions arise: Are LLMs actually able to effectively perform chemistry tasks? Or, Are they fundamentally limited for chemistry? In this paper, we demonstrate that our developed LLMs can achieve very strong results on a comprehensive set of chemistry tasks, outperforming the most advanced GPT-4 across all the tasks by a substantial margin and approaching the SoTA task-specific models. What makes such LLMs possible? First, we carefully construct a large-scale, comprehensive, and high-quality dataset for instruction tuning named SMolInstruct. We metic- ESOL (PP-ESOL) Query: How soluble is <SMILES> CC(C)Cl </SMILES> ? Response: Its log solubility is <NUMBER> -1.41 </NUMBER> mol/L. IUPAC to Molecular Formula (NC-I2F) Query: What is the molecular formula of the compound with this IUPAC name <IUPAC> 2,5-diphenyl-1,3-oxazole </IUPAC> ? Response: <MOLFORMULA> C15H11NO </MOLFORMULA> LIPO (PP-LIPO) Query: Predict the octanol/water distribution coefficient logD under the circumstance of pH 7.4 for <SMILES> NC(=O)C1=CC=CC=C1O </SMILES> . Response: <NUMBER> 1.090 </NUMBER> IUPAC to SMILES (NC-I2S) Query: Could you provide the SMILES for <IUPAC> 4-ethyl-4-methyloxolan-2-one </IUPAC> ? Response: Of course. It's <SMILES> CCC1(C)COC(=O)C1 </SM

b74899c4-d099-4761-81e9-cfe330fb5007

# Llasmol: Advancing Large Language Models For Chemistry With A Large-Scale, Comprehensive, High-Quality Instruction Tuning Dataset ## 1. Introduction the circumstance of pH 7.4 for <SMILES> NC(=O)C1=CC=CC=C1O </SMILES> . Response: <NUMBER> 1.090 </NUMBER> IUPAC to SMILES (NC-I2S) Query: Could you provide the SMILES for <IUPAC> 4-ethyl-4-methyloxolan-2-one </IUPAC> ? Response: Of course. It's <SMILES> CCC1(C)COC(=O)C1 </SMILES> BBBP (PP-BBBP) Query: Is blood-brain barrier permeability (BBBP) a property of <SMILES> CCNC(=O)/C=C/C1=CC=CC(Br)=C1 </SMILES>? Response: <BOOLEAN> Yes </BOOLEAN> SMILES to Molecular Formula (NC-S2F) Query: Given the SMILES representation <SMILES> S=P1(N(CCCl)CCCl)NCCCO1 </SMILES>, what would be its molecular formula? Response: It is <MOLFORMULA> C7H15Cl2N2OPS </MOLFORMULA> . ClinTox (PP-ClinTox) Query: Is <SMILES> COC[C@@H](NC(C)=O)C(=O)NCC1=CC=CC=C1 </SMILES> toxic? Response: <BOOLEAN> No </BOOLEAN> SMILES to IUPAC (NC-S2I) Query: Translate the given SMILES formula of a molecule <SMILES> CCC(C)C1CNCCCNC1 </SMILES> into its IUPAC name. Response: <IUPAC> 3-butan-2-yl-1,5-diazocane </IUPAC> HIV (PP-HIV) Query: Can <SMILES> CC1=CN(C2C=CCCC2O)C(=O)NC1=O </SMILES> serve as an inhibitor of HIV replication? Response

220f0c46-9dd7-4afe-a55f-79b9d63da59a

# Llasmol: Advancing Large Language Models For Chemistry With A Large-Scale, Comprehensive, High-Quality Instruction Tuning Dataset ## 1. Introduction given SMILES formula of a molecule <SMILES> CCC(C)C1CNCCCNC1 </SMILES> into its IUPAC name. Response: <IUPAC> 3-butan-2-yl-1,5-diazocane </IUPAC> HIV (PP-HIV) Query: Can <SMILES> CC1=CN(C2C=CCCC2O)C(=O)NC1=O </SMILES> serve as an inhibitor of HIV replication? Response: <BOOLEAN> No </BOOLEAN>