When developing software for materials R&D, one unavoidable reality stands out: while engineers naturally express themselves through language, computational tools require inputs defined by structures, compositions, and parameters. Historically, the burden of translating between these two "languages" has fallen on humans. With MatterSeek, our primary goal is to bridge this gap-allowing users to simply "state" their requirements, while the system handles the subsequent calculations.
Product Positioning: A Single Entry Point, Dual-Layer Capabilities
MatterSeek is an AI platform for materials R&D developed by Yuyao Technology. It utilizes natural language as its entry point, supported by two core layers: a Materials Science Large Language Model (LLM) and a proprietary Hierarchical Neural Network (HNN). The LLM interprets user intent, parses chemical elements and performance constraints, and extracts synthesis pathways and experimental conditions from scientific literature. The HNN, meanwhile, performs high-dimensional predictions and inverse design, mapping the relationships between composition, structure, processing, and performance.
We position the Materials Science LLM as an "interpreter and orchestrator" rather than a direct generator of numerical performance values. The reasoning is straightforward: in materials R&D, an incorrect figure-such as a wrong bandgap value or synthesis temperature-could lead to weeks of wasted experimental effort. Language models excel at reading literature, deconstructing complex problems, and translating vague requirements into precise, computable constraints; however, the actual numerical prediction is left to models grounded in physical principles. This is a fundamental design principle for our product.
The Language Layer Ensures Clarity; The Physics Layer Ensures Accuracy
Users can input a request such as, "Find a perovskite containing Sn and Pb with a bandgap between 1.5 and 2.0 eV." The Materials Science LLM first identifies the elements, parses the performance ranges, and retrieves relevant knowledge from literature, converting the natural language into a structured query. The HNN model then performs predictions based on a 1035-dimensional material descriptor system, outputting candidate compositions, crystal structures, and synthesis process parameters. The platform incorporates specialized sub-models covering inorganic, organic, and biomedical domains, designed to maintain high accuracy even with limited datasets-a critical requirement for real-world systems where experimental data is often scarce.
We do not believe that language models alone can "discover" new materials. Its value lies in connecting literature-based knowledge, user intent, and computational tools. This process narrows the search space, ensuring that experimental resources are reserved for candidates truly worthy of validation.
This closed-loop system is what primarily distinguishes us from approaches that rely solely on software.
MatterSeek is integrated with Yuyao's proprietary automated experimental platform. Predicted candidates can proceed directly to high-throughput synthesis and characterization, with the resulting empirical data fed back to refine the model. Knowledge extracted from literature by our Materials Science Large Language Model, performance requirements specified by users, and candidate solutions generated by our HNN (Hybrid Neural Network) are all subjected to rigorous "wet-lab" validation. The outcomes of these experiments then serve as training signals for the next round of predictions. Without this feedback loop, even the most accurate standalone model is merely a generator of candidates that "seem plausible." By successfully operationalizing this entire cycle, we have transformed our software from a simple "suggestion tool" into an integral component of the R&D pipeline.
We are clear about its boundaries.
MatterSeek is currently deployed across sectors such as aerospace, semiconductors, new energy, and eco-friendly, biodegradable materials. It serves dozens of enterprises and has supported over a hundred R&D projects. While these figures may be modest, they underscore a fundamental truth: the value of materials R&D software does not lie in replacing materials scientists, but in liberating human judgment from the drudgery of exhaustive trial-and-error and literature review-allowing experts to focus instead on defining problems, understanding underlying mechanisms, and making critical, high-level decisions.
In this framework, the Materials Science Large Language Model functions as a high-bandwidth knowledge interface and a central orchestrator-bridging the gap between "stating requirements" and "calculating materials." Ensuring the robustness of this connection is far more important than simply increasing the number of model parameters.

