Physically Consistent Super-Resolution Generative Adversarial Network for General Transient Simulations

Machine Learning – especially through Generative Adversarial Networks (GANs) – has brought transformative advancements to Super-Resolution (SR) techniques. Despite these improvements, the resulting images often lack physical relevance, which is crucial for scientific and engineering applications.

This research, led by a PhD student from Curtin University during a three-month DISIPA internship at CSIRO supported by the Discovery Program, is published in IEEE Transactions on Pattern Analysis and Machine Intelligence (https://doi.org/10.1109/TPAMI.2025.3596647). It introduces PC-SRGAN (Physically Consistent Super-Resolution Generative Adversarial Network), a novel framework designed to enhance image resolution from transient numerical simulations on coarse meshes, while preserving physical integrity for more interpretable simulations. This method demonstrates substantial gains in both Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) compared to traditional SR approaches – even when trained on a fraction of the data (e.g., just 13% of the dataset achieves performance comparable to SRGAN).

Beyond improving resolution, PC-SRGAN contributes to the broader field of physically grounded machine learning by integrating validated time-stepping algorithms and sophisticated evaluation metrics. Its emphasis on physical consistency makes it particularly suitable as a surrogate model for dynamic, time-dependent systems—setting it apart from conventional SR methods.