Convfeatnet Ensemble: Integrating Microstructure and Pre-Defined Features for Enhanced Prediction of Porous Material Properties

Abstract

This study introduces ConvFeatNet, a deep learning framework specifically designed to predict the mechanical properties of porous materials based on their microstructures. Despite dataset limitations, ConvFeatNet integrates both structural and predefined features with deep learning techniques to enhance predictive accuracy. The ensemble version of ConvFeatNet achieves a mean absolute error (MAE) of 0.85 J/m2 in predicting fracture energy using 1,000 samples, outperforming a simple MLP (MAE: 1.08 J/m2) and CNN (MAE: 1.38 J/m2) by 21% and 38%, respectively. Expanding the dataset to 10,000 samples further reduces the MAE to 0.51 J/m2, representing a 24% improvement over the MLP and a 9% improvement over the CNN. Additionally, SHAP analysis is employed to interpret model predictions, revealing the key structural determinants influencing mechanical behavior. This study highlights the synergy between deep learning and domain knowledge, offering a robust approach for deciphering the mechanical properties of porous materials.

Yuhai Li

Ph.D. student of Department of Civil and Environmental Engineering

My research interests include data-driven material analysis and machine/deep learning for material science.