Yuyao MatterSeek — Universal Foundation Model for New Materials Prediction (Beta)

MatterSeek

A Physics-Reasoning-Driven, Language-Agnostic Universal Foundation Model for New Materials Prediction

The End of Trial and Error — A New R&D Paradigm of Inverse Design and Precision Synthesis

Let materials evolve like living systems ｜ Accelerate the discovery of new materials ｜ Fewer experiments, better performance

Website · Online Demo (HF Space) · Technical Report / Paper

Introduction

MatterSeek is a physics-reasoning-driven universal foundation model for new materials prediction, developed by Yuyao Technology. Built on our proprietary HNN (Hierarchical Neural Network) architecture, it serves inorganic and organic materials R&D as a one-stop AI materials development tool for researchers worldwide.

MatterSeek breaks through the strong coupling of composition–structure–processing–property in high-dimensional space, enabling accurate forward and inverse design with precise prediction, and closes the loop through wet-lab validation on Yuyao's in-house experimental platform. MatterSeek is not just a prediction tool — it is an industrially deployable, high-confidence execution solution.

Key Features

Advanced algorithms with physical constraints: physics-reasoning-driven, with 92% of parameters self-developed
1,035 material descriptors + 188 domain-specific sub-models: precisely adapted to different material systems
Broad coverage: inorganic and organic materials R&D
Closed experimental loop: 7×24 automated experiments in our own laboratory; experimental data feeds back into the model for recursive training and continuous improvement

Core Capabilities

Capability	Description
Small-sample prediction	Reliable predictions from just tens of data points
Natural-language chemistry instructions	Describe your chemistry requirements in plain language — no prompt engineering needed
Elemental constraint filtering	Quickly narrow the candidate material space by elemental composition ranges
Custom property targets	Flexibly define target properties for on-demand forward design
Multi-dimensional visualization	Inspect prediction results from multiple perspectives to support decision-making

How It Works: The Experimental Closed Loop

Our materials R&D operates as a self-evolving closed loop, powered by the MatterSeek predictive model. The process flows through five connected stages:

Stage 1 – AI-Driven Design. Target properties are decomposed into quantifiable metrics, and MatterSeek predicts candidate formulations and process parameters. Virtual screening narrows these down to three candidate solutions.

Stage 2 – Automated Experiment Validation. Process instructions are dispatched automatically to a robotic synthesis pipeline, where samples are prepared, characterized across multiple dimensions, and data is captured in structured form — running 24 hours, unmanned.

Stage 3 – Data Benchmarking & Decision. Measured values are automatically compared against predictions. A deviation below 10% is marked qualified; anything above triggers iteration.

Stage 4 – Materials Database Deposit (qualified path). Every qualified formulation, along with its process and characterization data, is structured into our proprietary "Material–Process–Property" knowledge base — turning each success into a permanent, reusable asset.

Stage 5 – Model Self-Evolution (unqualified path). Failed predictions are not discarded but fed back: active learning annotates the samples, genetic algorithms re-screen the candidate population, and the model's parameters and feature space are reconstructed.

The loop then returns to Stage 1 with a sharper model. Because the deviation between prediction and experiment is itself the learning signal, each round compounds: qualified results build the data moat, unqualified ones upgrade the engine. Within three feedback rounds we converge on the target — delivering exponential gains in R&D efficiency and a competitive data asset that grows with every experiment.

Video Demos

Promo Video (60s)

Product Walkthrough (3min)

Suggested demo path: Sign in → enter a natural-language instruction or target properties → filter by elemental constraints → explore multi-dimensional result visualization → export composition / processing parameters → submit for offline validation

Model Architecture

Item	Description
Architecture	Proprietary HNN hierarchical neural network (physics-reasoning-driven). Two-tier design: a shared backbone learns cross-domain material representations from the 1,035-d fingerprint; on top sit 192 second-level expert sub-models grouped under 17 first-level vertical domains, each fine-tuned on materials that exhibit the target property. Physics-informed regularization embeds symmetry, conservation laws and thermodynamic constraints as soft penalties, so the network outputs physically plausible predictions rather than pure interpolation.
Parameters	≈ 0.5 B total: ~80 M shared backbone + 192 lightweight domain adapters (~2 M each).
Material descriptors	1,035
Domain sub-models	192 second-level (二级) performance-variable sub-models organized under 17 first-level (一级) vertical domains: superconductivity, magnetocaloric, magneto-transport, magnetism, catalysis & surfaces, electronic transport, optics, nuclear & irradiation, mechanics, crystal structure, thermoelectrics, thermodynamics, thermal transport, phonons, ferroelectrics, barocaloric, and fabrication & cost. Each sub-model is a dedicated predictor fine-tuned on materials that exhibit the target property. See category breakdown below.
Input formats	Natural-language instructions, property targets, composition formulas.
Output formats	Predicted properties with uncertainty bounds, recommended compositions (Pareto-front candidates), processing parameters (sintering / annealing windows), and ranked experiment proposals.
Training	Two-stage curriculum: (1) Pre-training on large-scale composition–structure–property records via self-supervised masked-fingerprint / structure reconstruction; (2) Domain fine-tuning with physics-informed loss, active-learning feedback loops, and Bayesian / NSGA-guided candidate selection.

Domain Sub-model Breakdown (17 first-level, 192 second-level variables)

Category	#	Representative Second-level Variables
Superconductivity	12	Tc (conventional / cuprate / Fe-based / heavy-fermion), Hc1 / Hc2 / Hirr, critical current density Jc, coherence length ξ, superconducting gap Δ, optimal carrier concentration
Magnetocaloric	4	Magnetocaloric ΔSm, ΔTad, refrigerant capacity RC, magnetocaloric strain
Magneto-transport	7	Hall coefficient, magnetoresistance, anomalous Hall (AHE), spin-phonon / spin-spin / spin-orbit scattering, spin relaxation time
Magnetism	19	Curie / Néel temperature, saturation magnetization, remanence, coercivity Hc, susceptibility, permeability, magnetic moment / spin moment / paramagnetic moment, anisotropy K, energy product BHmax, magnetostriction, exchange coupling J
Catalysis & Surfaces	13	OER Tafel slope / onset / limiting current, exchange current j0, ionic conductivity, photocatalytic activity, BET surface area, zeta potential, surface energy, corrosion Ecorr/icorr, biocompatibility (hemolysis)
Electronic Transport	22	Conductivity, electron/hole mobility, carrier concentration, effective mass, direct/indirect band gap, DOS, work function, residual resistivity RRR, e-e scattering, dielectric loss tanδ, field dependence of resistivity, charge neutrality check
Optics	18	Permittivity ε′, dielectric loss ε″, refractive index, birefringence, transmittance / extinction, reflectivity, optical absorption, optical conductivity, plasma frequency, luminescence wavelength / lifetime / PLQY, optical phonon TO/LO, LIDT, Abbe V, nonlinear χ⁽²⁾, breakdown strength
Nuclear & Irradiation	2	Neutron absorption cross-section, radiation stability
Mechanics	26	Elastic stiffness / compliance, Poisson ratio, bulk / shear / Young modulus, yield / tensile / compressive / flexural / fracture strength, hardness, fracture toughness KIC, fatigue, creep, thermal shock, residual stress, density ρ, sintering shrinkage, hydrogen storage (capacity / PCT / kinetics / cycling)
Crystal Structure	8	7 crystal systems, lattice constant a/b/c, structural phase transition temperature, bulk modulus, linear thermal expansion CTE, volume change (at transition)
Thermoelectrics	3	Seebeck coefficient S, power factor PF, figure of merit ZT
Thermodynamics	22	Tm, Tb, Tg, enthalpy / entropy of formation, reaction enthalpy / entropy, enthalpy of fusion / vaporization, decomposition temperature, heat capacity, convex hull energy / hull distance, critical cooling rate Rc, critical nucleation size / TTT nose
Thermal Transport	8	Thermal conductivity (phononic / electronic), thermal diffusivity α, Lorenz number L, electronic heat capacity, thermal expansion, thermal emissivity
Phonons	10	Longitudinal / transverse sound velocity, acoustic impedance, shear / Young modulus, Debye temperature, shear / Young loss factors
Ferroelectrics	13	Spontaneous / remanent polarization, coercive field Ec, piezoelectric d, pyroelectric p, electromechanical coupling k, mechanical Q, magnetoelectric αME, ferroelectric Tc
Barocaloric	4	Barocaloric ΔSm, ΔTad, refrigerant capacity, barocaloric strain
Fabrication & Cost	1	Fabrication process route and manufacturing cost

Training Data

Dimension	Description
Data sources	Predominantly proprietary high-throughput experimental databases and curated literature measurements (property values); structural data (crystal structure, lattice, space group) supplemented by DFT calculation.
Scale	Several million composition–structure–property records, experiment-dominant (experimental property measurements paired with computed or measured structures), continuously expanding via active learning.
Material systems	Material systems are selected dynamically according to the target performance variable (each sub-model is trained on the materials that exhibit that property), rather than restricted to a fixed alloy / ceramic list.
Cleaning & deduplication	Automated pipeline: unit-cell normalization, phase / space-group reconciliation, composition-rounding canonicalization, duplicate-record merging, outlier detection against DFT↔experimental consistency, and train/val/test split by material family to prevent leakage.

Benchmarks

Accuracy is reported on held-out experimental test sets split by material family within our own corpus. No public benchmark or peer model covers this exact set of 192 performance variables, so cross-model peer comparison is not applicable.

Quick Start

from [library] import [ModelClass]
model = [ModelClass].from_pretrained("[org]/matterseek")
result = model.predict([input: natural-language instruction / target properties / composition])
print(result)  # minimal runnable example to be added upon model release

Need API access or offline experimental validation? We provide API services and a closed validation loop (forward-designed compositions → laboratory synthesis → characterization feedback). Visit our website to get in touch.

Use Cases

Alloy composition optimization: rapidly screen candidate composition ranges given service-performance requirements
High-entropy alloy R&D: explore multi-principal-element composition spaces, outputting compositions + processing windows
New materials discovery: inverse design of novel compositions with crystal-structure prediction
Processing parameter recommendation: melting / sintering / heat-treatment parameters based on target microstructure and properties
[TODO: organic / biopharma use cases]

Experimental Platform

MatterSeek's closed loop is backed by our in-house laboratory, with end-to-end capabilities from synthesis to characterization:

Coverage & Best Practices

Coverage Boundary MatterSeek delivers production-grade predictions within its 17 vertical domains and 192 target properties. For emerging element combinations or properties outside the current taxonomy, the model can be extended on demand via our active-learning loop — a few new experiments bootstrap a new sub-model rather than a full retraining. Recommended Use ① Use predictions as a fast screening and decision-support layer to prioritize experiments, not as a substitute for final characterization. ② Pair model output with the recommended composition–processing window; the Pareto-front candidates are ranked for multi-objective design. ③ For critical or regulated applications, validate top candidates offline before scale-up. ④ The model is actively expanding into additional verticals (e.g., biopharma toxicity, long-term reliability) — contact us for the latest roadmap.

License

This model is released under the CC BY-NC 4.0 (Attribution-NonCommercial 4.0 International) license.

Research use: permitted — free to use, modify, and share for academic research, teaching, and personal non-commercial purposes, with attribution
Commercial use: license required — we offer API services, on-premise deployment, and closed-loop experimental validation. Visit our website to contact us

Citation

If MatterSeek helps your research, please cite:

@misc{matterseek2026,
  title  = {MatterSeek: [TODO: official paper title]},
  author = {[TODO: authors]},
  year   = {2026},
  url    = {https://huggingface.co/[org]/matterseek}
}

Yuyao Technology (Beijing) Co., Ltd. · MatterSeek Team

Website · Contact Us

Downloads last month: -; Downloads are not tracked for this model. How to track