Physics-informed generative adversarial networks with explainable AI for battery cooling plate design: A convergence of physics, machine learning, and thermal engineering

Abstract

Liquid-cooled plates are essential for EV battery thermal management, but conventional design relies on expensive topology optimization or iterative CFD. Generative Adversarial Networks (GANs) offer a faster alternative yet suffer from two key limitations: no physical constraint enforcement during generation, and black-box operation with no design rationale.

This work presents a Physics-Informed GAN (PI-GAN) integrating adversarial learning, multi-term physics regularization, and explainable AI (XAI). Trained on 81 topology-optimized cooling plate designs, PI-GAN incorporates seven physics losses (continuity, smoothness, connectivity, area ratio, envelope, spectral fidelity, binarization), InfoGAN-structured latent codes for controllable generation, and an XAI module (PCA, tSNE, latent traversal, Q-MIG).

PI-GAN achieves SSIM of 0.4046 ± 0.0126 (1.7× baseline GAN) and reduces edge complexity by 2.5×. Unlike the baseline GAN which exhibits near-total mode collapse (intra-SSIM = 0.998), PI-GAN produces diverse designs (intra-SSIM = 0.445). The XAI module identifies which latent dimensions control specific geometric features, enabling interpretable generation.

The convergence of physics regularization, adversarial learning, and XAI yields generative models that are physically valid, diverse, and interpretable extendable to bone scaffolds and metamaterials.