DreamLufei/invDesMobility-diffcsp-generator

InvDesMobility DiffCSP Generator Checkpoints

This repository contains the DiffCSP generator checkpoints used by the InvDesMobility inverse-design workflow. The checkpoints are released as external artifacts for the code repositories:

• https://github.com/DreamLufei/invDesMobility
• https://github.com/DreamLufei/invdesmobility_loop
• https://github.com/DreamLufei/2d-mobility

Paper and Repositories

• Paper: InvDesMobility: a reliability-gated first-principles feedback framework for closed-loop materials discovery
• arXiv: https://arxiv.org/abs/2606.16133
• Project page: https://dreamlufei.github.io/invDesMobility/
• GitHub: https://github.com/DreamLufei/invDesMobility
• Loop repository: https://github.com/DreamLufei/invdesmobility_loop
• Mobility workflow: https://github.com/DreamLufei/2d-mobility
• Zenodo: https://doi.org/10.5281/zenodo.20475023

Files

• pretrained/PretrainGenerationModel.ckpt: upstream DiffCSP warm-start checkpoint used by the pipeline.
• finetuned/mobility2d_highquality280_ft_v1/best.ckpt: seed mobility-2D fine-tuned generator.
• finetuned/generator_round_XX/best.ckpt: closed-loop feedback fine-tuned generators for rounds 01, 02, 03, 04, 06, 07, 08, and 09.
• Each fine-tuned directory also includes hparams.yaml, lattice_scaler.pt, and prop_scaler.pt.

The epoch=...ckpt files from local training logs are intentionally omitted because they are byte-identical to the retained best.ckpt files for the corresponding rounds. Run logs, W&B files, generated pools, and VASP outputs are not included.

Intended Use

These checkpoints are intended for reproducing the candidate-generation stage of InvDesMobility and for research use in feedback-guided 2D crystal generation.

Training Data

The seed model was fine-tuned on the mobility-2D high-quality seed dataset. Closed-loop checkpoints were fine-tuned on feedback-augmented DiffCSP datasets constructed from trusted first-principles mobility-validation records.

The matching datasets are packaged separately in DreamLufei/invDesMobility-data.

Training Parameters

The retained hparams.yaml files contain the exact DiffCSP/PyTorch Lightning configuration for each checkpoint, including model architecture, dataset path, batch size, optimization configuration, and diffusion scheduler settings.

Key settings used by the feedback models include:

• DiffCSP CSPDiffusion with CSPNet decoder.
• hidden_dim=512, num_layers=6, max_neighbors=20, radius=7.0.
• Fine-tuning via the project scripts under 05_steps/02_finetune_generator/ and 05_steps/09_closed_loop_feedback/.

Evaluation

These generator checkpoints were evaluated operationally inside the closed-loop InvDesMobility workflow: structures generated from each checkpoint were deduplicated, screened by surrogate models, and then selected candidates were validated with first-principles mobility calculations.

The generated pools, feedback datasets, and retained-channel records used to audit this process are packaged in DreamLufei/invDesMobility-data. No standalone generative benchmark table is included in this model repository, because the relevant quality measure for this work is downstream retention and DFT validation rather than raw sample likelihood alone.

Limitations

The generator proposes candidate structures; it does not validate mobility, dynamic stability, synthesizability, or DFT-level electronic structure. Candidate structures require downstream deduplication, surrogate screening, and first-principles validation with the companion 2d-mobility workflow.

Citation

If you use these checkpoints, please cite the InvDesMobility manuscript and the associated GitHub repositories above.