Wavelet scattering networks for atomistic systems with extrapolation of material properties

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Machine learning in materials science aims to predict material properties beyond interpolation.
• Developing feature representations that preserve physical symmetries is crucial for accurate predictions.
• Atomic orbital wavelet scattering transform (AOWST) is a successful featurization method for machine learning energy prediction.

Purpose of the Study:

• To test the generalizability of machine learning models trained with AOWST features.
• To evaluate the prediction of properties not included in the original training set, such as elastic constants and migration barriers.
• To assess the effectiveness of statistical feature selection in improving extrapolation accuracy.

Main Methods:

• Utilizing the atomic orbital wavelet scattering transform (AOWST) for feature representation of 3D atomic systems.
• Training machine learning models on energy prediction tasks for small molecules and the amorphous LiαSi system.
• Applying statistical feature selection methods to identify the most relevant features for extrapolation.
• Evaluating model performance on predicting elastic constants and migration barriers for LiαSi.

Main Results:

• Machine learning models with AOWST features achieved accuracy comparable to density functional theory for energy prediction.
• The models demonstrated remarkable accuracy in extrapolating to predict properties like elastic constants and migration barriers.
• Statistical feature selection effectively reduced overfitting and enhanced predictive accuracy for unseen properties.

Conclusions:

• The AOWST featurization method enables machine learning models to generalize beyond training data.
• Accurate prediction of diverse material properties, including elastic constants and migration barriers, is achievable through extrapolation.
• Statistical feature selection is a key technique for improving the reliability and accuracy of machine learning models in materials science.