MatterSeek

MatterSeek is a physics-reasoning universal foundation model for new-materials prediction (Beta) from China’s Yuyao Technology. Hugging Face model card: https://huggingface.co/YuyaoTechnology/MatterSeek ; product site: https://yuyaomatterseek.net .

Positioning: inorganic and organic materials R&D via forward/inverse design across the coupled composition–structure–process–property space, closed with an in-house wet-lab loop.

Highlights:

• Vendor claims ~92% self-developed algorithms; 1,035 material descriptors + 192 domain sub-models under 17 first-level verticals;
• Core features: reverse design (target properties → composition/process/structure), element-constraint filtering, custom property metrics, and multi-dimensional visualization;
• In-house lab runs automated experiments 24/7 with data fed back for recursive training;
• HF model card shows about ~2,888 likes; related Space newly entered Hugging Face Spaces trending (per radar excerpt).

Problem

New-materials discovery still burns many trial-and-error experiments. MatterSeek aims to screen and inverse-design under physics constraints with a hierarchical network, then validate on a proprietary experimental platform that continually refreshes the model.

Architecture & data (vendor description)

• Proprietary HNN hierarchical neural net: a shared backbone learns cross-domain fingerprints; expert sub-models sit on top by vertical; physics-informed regularization soft-penalizes symmetry, conservation, and thermodynamic violations;
• Example first-level domains: superconductivity, magnetocaloric/barocaloric, magneto-transport, magnetism, catalysis & surfaces, electronic transport, optics, nuclear & irradiation, mechanics, crystal structure, thermoelectrics, thermodynamics, thermal transport, phonons, ferroelectrics, fabrication & cost;
• Training data mostly proprietary high-throughput experiments and curated literature properties, with structures supplemented by DFT; scale described as several million composition–structure–property records;
• Accuracy reported on held-out experimental sets split by material family; vendor says no public benchmark covers all 192 performance variables, so peer cross-model comparison is not claimed.

Limits

• Beta; benchmarks and scale figures are largely vendor-reported;
• Intended as a screening/decision layer, not a substitute for final characterization; regulated uses need offline validation;
• Outside the current 17×192 coverage, new element/property combos need active-learning extensions.

Who it's for

• Inorganic/organic materials R&D, high-throughput experiment, and materials-informatics teams;
• Labs ranking composition/process candidates from target properties;
• Researchers watching physics-constrained materials foundation models with lab feedback loops.

Getting started

1. Read https://huggingface.co/YuyaoTechnology/MatterSeek and visit https://yuyaomatterseek.net ;
2. Suggested path after sign-in: natural-language or target properties → elemental constraints → multi-view results → export composition/process params → offline validation;
3. Treat outputs as a triage layer; characterize Pareto-front candidates (e.g., XRD/SEM) before scale-up;
4. For properties outside the current taxonomy, ask about active-learning sub-model extension rather than assuming full coverage.