Physics-Informed Machine Learning Unlocks Data-Efficient Pathways for Materials Characterization

A research team from KAIST led by Professor Seunghwa Ryu has demonstrated that material properties can be reliably identified from only a minimal amount of experimental data by embedding physics-informed machine learning (PIML) into the modeling process. Their two recent studies, published in Computer Methods in Applied Mechanics and Engineering (CMAME) and npj Computational Materials, highlight how coupling physical laws with neural networks enables breakthroughs in both mechanical and thermoelectric materials research.

In the CMAME study, the team addressed the challenge of characterizing hyperelastic materials such as rubber. By incorporating governing equations directly into a PINN-based framework, they showed that full-field displacement fields and constitutive equations could be reconstructed from a single experiment using only displacement and reaction-force data. Unlike conventional methods requiring extensive stress–strain datasets, the proposed approach delivered robust predictions even under noisy or data-scarce conditions. This work was conducted in collaboration with Professor Jae-Hyuk Lim of Kyung Hee University, with PhD students Hyeonbin Moon and Donggeun Park as co-first authors.

In the npj Computational Materials article, the researchers extended the PIML approach to thermoelectric materials. They developed an inverse design framework capable of estimating temperature-dependent thermal conductivity and Seebeck coefficients from only a few voltage and temperature measurements. The introduction of a Physics-Informed Neural Operator (PINO) further allowed the model to generalize across previously unseen materials without retraining. Trained on data from only 20 materials, the framework successfully predicted the properties of 60 new materials with high accuracy, underscoring its potential for rapid and large-scale screening of materials. This work was carried out with Dr. Byeongki Ryu of the Korea Electrotechnology Research Institute (KERI), with PhD students Hyeonbin Moon, Songho Lee, and Wabi Demeke serving as co-first authors.

“These studies show that by embedding physical laws into machine learning models, we can achieve both data efficiency and physical consistency,” said Professor Ryu. “This marks a critical step toward accelerating the discovery and validation of next-generation materials.”

Both projects were supported by the InnoCORE Program, the National Research Foundation of Korea, and the Ministry of Science and ICT, with additional support from the Ministry of Food and Drug Safety for the CMAME study. Experts anticipate that these advances will significantly reduce experimental costs while opening new pathways for AI-driven materials innovation.