Physics-grounded AI framework aims to make predictions about new materials more testable

A new framework, dubbed Physics-Grounded Materials AI (PhysMat AI), has been proposed to enhance the reliability of materials discovery through the integration of fundamental physical laws into artificial intelligence systems. The journal Advanced Functional Materials recently published the research, which introduces a five-part framework that includes roles such as prior knowledge, descriptors, constraints, verifiers, and infrastructure.

These roles guide the representation of materials data, AI reasoning, and the evaluation of predictions against physical principles.

Hao Li, a distinguished professor at the Advanced Institute for Materials Research at Tohoku University, emphasizes that materials discovery cannot rely solely on data correlations. Instead, incorporating physical principles into AI can make predictions more interpretable, testable, and meaningful from a materials science perspective. The five roles specified in the framework work together to guide the development and evaluation of AI-generated predictions.

By applying physical knowledge to AI systems, the researchers argue that materials discovery can move beyond correlation-based methods toward reasoning based on physical principles. Examples of its application include catalysis, solid-state electrolytes for solid-state batteries, and hydrogen storage materials. In these areas, physical principles can help define search spaces, evaluate predictions, and connect AI-generated results with experimental validation.

The researchers propose a development pathway from physics-aware AI to physics-reasoning AI and ultimately to physics-autonomous AI. This progression could support materials discovery for energy technologies by enabling AI to integrate physical reasoning, simulations, and experimental data in a continuous discovery process. The framework aims to bridge the gap between AI-based predictions and established principles of materials science, making predictions more interpretable and experimentally testable.