Evaluating the use of a machine learning simulator for structure-property prediction: A case study on disordered...

Area of Science:

• Materials Science
• Computational Physics
• Machine Learning

Background:

• Machine learning models often require extensive datasets, limiting their application in scientific discovery where data generation is challenging.
• Generalizing beyond training data is crucial for discovering novel materials and solutions in engineering.

Purpose of the Study:

• To investigate the use of dynamical data and graph neural networks for efficient system-to-property learning.
• To enable accurate out-of-distribution prediction for materials behavior under mechanical stress.

Main Methods:

• Developed a graph neural network-based simulator utilizing dynamical data.
• Trained the simulator on uniaxial compression of 2D disordered elastic networks.
• Evaluated the model's ability to predict emergent properties and generalize to unseen conditions.

Main Results:

• The simulator learned underlying physical dynamics from limited data and accurately predicted temporal evolution.
• Emergent properties like Poisson's ratio were predicted accurately without explicit training.
• The model demonstrated strong generalization across temperature, strain amplitude, and out-of-distribution Poisson's ratios.

Conclusions:

• Dynamical data enhances information efficiency and generalizability in machine learning for materials design.
• This approach is particularly beneficial in data-scarce settings for scientific and engineering applications.
• Graph neural network simulators show promise for accelerating the discovery of novel materials with desired properties.