Imbalance-Aware Deep Surrogate Modeling in Structural and Thermal Response Prediction

Deep learning has shown significant promise in surrogate modeling for engineering simulations. However, when applied to structural and thermal response prediction, deep neural networks often suffer from performance degradation due to data imbalance—where low-magnitude responses dominate, and critical high-response regions are underrepresented. To address this challenge, we propose an imbalance-aware deep surrogate modeling framework that integrates convolutional neural networks (CNNs) with two imbalance mitigation strategies: stratified sampling and label transformation. The framework is applied across three representative case studies: (1) stress prediction in 2D cantilever beams, (2) transient thermal and thermal stress responses in heat-transfer fins, and (3) 3D stress prediction in cantilever beams under variable loading. We demonstrate that stratified sampling improves performance on small datasets by balancing rare and frequent response values, while label transformations reshape skewed label distributions to reduce model bias. Our results show consistent improvements in accuracy, particularly in predicting maximum response values, which are critical for design and safety applications. The proposed framework offers a scalable and adaptable approach for improving surrogate modeling in structural and thermal analysis and can be extended to more advanced architecture and complex simulations.