Artificial intelligence (AI) has become a popular tool for discovering new materials, but conventional machine learning approaches face important limitations. They often cannot explain why they make specific predictions, fail when applied to situations outside their training data, and sometimes violate well-established physical laws. A new framework called Physics-Grounded Materials AI (PhysMat AI) aims to solve these problems by integrating physics directly into the AI workflow.
The framework was developed by researchers at Tohoku University in Japan and described in a recent Advanced Functional Materials perspective paper. The researchers explain that materials discovery needs to move beyond simply finding correlations in data. By incorporating physical principles, AI predictions become more interpretable, testable, and meaningful for materials science.
How it's organized
The PhysMat AI framework divides physical knowledge into five complementary roles that guide the entire discovery process. Prior knowledge provides the scientific foundation. Descriptors define how material data are represented. Constraints limit the search space to physically realistic possibilities. Verifiers check predictions against physical principles. Infrastructure supports data management and model integration. Together these layers create what the authors call a closed-loop workflow, where experimental results feed back continuously to improve both the knowledge base and the AI models.
The researchers illustrate their framework with examples from practical applications: catalysts for speeding up chemical reactions, solid-state electrolytes for safer batteries, and materials for storing hydrogen fuel. In each case, knowing the underlying physics helps guide AI searches toward promising candidates and away from impossible ones. The authors also propose AI agents that could combine physics-aware components with databases, simulations, and experiments to formulate hypotheses and assess physical feasibility automatically.
The development pathway they envision moves from today's physics-aware AI, which merely incorporates physical knowledge, through physics-reasoning AI that actively uses such knowledge during scientific reasoning, to eventual physics-autonomous AI capable of integrating physical reasoning with simulations and experiments in a continuous discovery process. Several challenges remain, including deploying these systems at scale, developing models that truly reason with physics, and building appropriate data infrastructure.
This work matters because it offers a way to make AI-based materials discovery more trustworthy. By connecting machine learning predictions to established scientific principles, researchers can construct explanations that can be tested experimentally. For energy technologies in particular - better catalysts, safer batteries, more efficient hydrogen storage - this approach could accelerate practical innovation while maintaining scientific rigor.