TY  - JOUR
AU  - Gautam, Junesh
AU  - Acharya, Praneel
AU  - Hu, Zhong
PY  - 2025
DA  - 2025/10/24
TI  - Deep Learning Approaches for Predicting Strain Energy in Heterogeneous Materials
JO  - Recent Progress in Materials
SP  - 016
VL  - 07
IS  - 04
AB  - This paper investigates data-driven strain energy prediction for heterogeneous elastic materials using the Mechanical MNIST benchmark dataset (60,000 28 × 28 stiffness maps by precomputed finite-element results). The results by classical regression models (linear regression, random forest, and gradient boosting) using hand-crafted features were compared with the results by the deep learning models (Convolutional neural network (CNN) and residual network (ResNet)) trained end-to-end on images. In held-out tests, CNN and ResNet achieve MSE ≈ 4.21–4.33 with R2 ≈ 0.982, substantially outperforming classical methods (MSE ≈ 15.8–50.0; R2 ≈ 0.79–0.93). In addition to accuracy, the train/inference cost and model limitations (dataset scope, loading modes, physics fidelity) were discussed. The model interpretability (saliency/occlusion) was analyzed to link learned features to stiffness patterns. These results support the use of images as energy surrogates as fast approximations for design and optimization, while outlining steps for real-material validation (e.g., in coupling with acoustic emission sensing and physics‐informed learning).
SN  - 2689-5846
UR  - https://doi.org/10.21926/rpm.2504016
DO  - 10.21926/rpm.2504016
ID  - Gautam2025
ER  -