Researchers at KAIST have developed a new way to identify material properties, which are traits like strength or heat conduction that define how substances behave. This step usually needs lots of experiments and costly tools, slowing down progress. The new approach mixes physical laws with artificial intelligence (AI). It uses Physics-Informed Machine Learning, called PIML, where AI learns by including physical laws directly in its process. This lets scientists explore new materials quickly, even with scarce data, and supports faster design in areas like mechanics, energy, and electronics.
The work involves collaboration with Kyung Hee University, and Korea Electrotechnology Research Institute. They showed how this method accurately finds properties with limited information.
Advances in specific material types
In one study, the researchers focused on hyperelastic materials, like rubber that stretches and returns to shape. They used a Physics-Informed Neural Network, or PINN, an AI model that follows physics rules, to pinpoint deformation behavior and properties from just one experiment's small data set. Unlike old methods needing huge, clean data, this works with noisy or incomplete info.
In another study, the researchers applied the new method to thermoelectric materials, substances that turn heat into electricity or electricity into heat. Using a PINN-based technique, they estimated key measures like thermal conductivity, how well heat moves through, and the Seebeck coefficient, a gauge of electricity generation efficiency, from few measurements.
The researchers also introduced a Physics-Informed Neural Operator, or PINO, an AI that grasps natural laws and applies to new materials without retraining. After training on 20 materials, it predicted properties for 60 unseen ones accurately. This could enable fast screening of many material candidates, cutting experiment needs while keeping results reliable.
The researchers noted this as the first real-world use of physics-aware AI in materials, expanding to various fields. They described the methods and results of this study in two papers (1, 2) published in Computer Methods in Applied Mechanics and Engineering and npj Computational Materials.