Building Chemical Property Models for Energetic Materials from Small Datasets Using a Transfer Learning Approach.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Predicting chemical properties like impact sensitivity is challenging due to limitations in physics-based models and small experimental datasets.
• Existing methods struggle with extrapolation to new molecules and lack sufficient data for traditional machine learning (ML).

Purpose of the Study:

• To develop and demonstrate a novel transfer learning approach for accurately predicting experimentally measured chemical properties.
• To overcome the limitations of small experimental datasets and poorly extrapolating physics-based models in chemical property prediction.

Main Methods:

• A multi-target regression model was trained using transfer learning, combining a small set of experimental data with a large set of computed properties.
• A directed message-passing neural network (D-MPNN) ML model architecture was employed.
• The methodology was applied to predict the impact sensitivity of energetic crystals.

Main Results:

• The transfer learning approach, utilizing a D-MPNN, significantly improved prediction accuracy for experimentally measured properties.
• Both the characteristics of the computed dataset and the ML model architecture were found to be critical for prediction performance.
• The D-MPNN model with transfer learning outperformed both direct ML and physics-based models on a diverse test set.

Conclusions:

• Transfer learning offers a powerful strategy to enhance ML model performance for chemical property prediction, especially with limited experimental data.
• The developed D-MPNN model demonstrates broad applicability for various structure-property relationship modeling tasks.
• This work provides a scalable and accurate method for predicting challenging chemical properties, advancing computational materials science.