Physics-Aware AI Indispensable for Materials Research: Performance Gaps Highlighted in Thermal Conductivity Prediction

September 13, 2026

Columbia Engineering USA

Overview

Columbia Engineering emphasizes the critical need for physics-aware AI in materials science, particularly for machine learning interatomic potentials (MLPs) that predict macroscopic properties from quantum-level structures. A new method developed by Michele Simoncelli’s group leverages fundamental physics to systematically assess ML models, revealing performance gaps in models lacking strong physical constraints when predicting properties like thermal conductivity. The Matbench Discovery leaderboard now incorporates a new test set and metric design to diversify evaluation, exposing weaknesses in models that may appear similar on crystal stability.

In Depth

Key Findings

Researchers at Columbia Engineering advocate that ‘physics-aware’ AI is paramount in the application of artificial intelligence (AI) to materials science, particularly for machine learning interatomic potentials (MLPs) that predict macroscopic properties from quantum-level structures. Their newly developed evaluation method has revealed significant performance gaps in models that do not strongly incorporate physical constraints when predicting critical thermomechanical properties like thermal conductivity.

Technical / Clinical Details

Professor Michele Simoncelli’s research group has developed a novel benchmark that systematically integrates fundamental physical principles into the evaluation of machine learning models. This benchmark focuses not only on whether MLPs accurately predict crystal stability, but also on how precisely they can reproduce thermodynamic and mechanical properties such as thermal conductivity, elastic modulus, and sound velocity. While conventional MLPs excelled at predicting the stability of structures present in their training datasets, their performance tended to degrade when predicting properties associated with out-of-distribution structures or dynamic phenomena like thermal vibrations. The research demonstrated that incorporating robust physical constraints into models (e.g., conservation laws, symmetries) significantly enhances the robustness and generalization capability of such property predictions. Furthermore, the Matbench Discovery leaderboard now features a new test set and metric design to diversify evaluation, revealing actual weaknesses in models that previously appeared similar based on crystal stability alone, particularly in more detailed physical property predictions.

Background & Context

Materials informatics is a powerful tool for accelerating the design and discovery of new materials, but it faces a fundamental challenge: how accurately do AI models reflect physical reality? MLPs, in particular, are key to enabling large-scale simulations, such as molecular dynamics, by maintaining the accuracy of first-principles calculations (DFT) while drastically reducing computational costs. However, purely ‘data-driven’ approaches with weak physical awareness risked ‘hallucinations’ (physically unreasonable predictions) or unexpected behaviors. This research highlights the urgency and importance of integrating a deep understanding of physics into AI model design at the interface of AI and materials science. This is an essential step towards establishing more reliable AI-driven material design processes and accelerating scientific discovery.

Strategic Significance & Outlook

The recognition of the need for physics-aware AI will significantly influence the direction of future AI research in materials science. As MLPs incorporate stricter physical constraints and better integrate with first-principles calculations, the development of various high-performance materials—including thermal management materials, batteries, superconductors, and semiconductors—is expected to accelerate. The introduction of new benchmarks and evaluation metrics will foster competition in developing more robust and reliable AI models, enhancing comparability and transparency across the research community. Ultimately, this approach will form the foundation for ensuring that AI-proposed materials are functional and possess industrially applicable properties, playing a crucial role in bridging the ‘lab-to-fab’ gap in materials development.

Source: https://www.engineering.columbia.edu/about/news/ai-materials-needs-be-more-physics-aware

Get our weekly technology intelligence — free

Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.

Subscribe Free — Weekly Tech Intelligence

By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.