Structurally Causal Data Augmentation via Physics-Informed Neural Networks

Abstract

The deployment of machine learning (ML) models in process engineering domain is limited by the availability of empirical failure data. This has caused any ML deployed in this domain to underperform, thus unable to generalize well to outof-distribution (OOD) condition. Several methods are available to augment and generate these synthetic failure scenarios, but some of these methods do not observe the first principles and physics law crucial to the domain. This paper proposes a causally validated data augmentation method utilizing Physics Informed Neural Network (PINNs) to generate the transient edge cases including the failure scenario data. It is empirically verified that the PINN-generated synthetic datasets preserve the structural causality inherent in the governing physical equations. Furthermore, to verify the necessity of the augmented dataset, it is trained on a Liquid Neural Network (LNN) with extensive ablation studies. This is to valuate the impact of the causallyembedded synthetic data on the LNN's training dynamics, model performance, and OOD generalization. The results demonstrate that enforcing structural causality during augmentation yields over a 41% improvement in anomaly variance capture, proving it is necessary for the LNN to accurately learned multi-physics failure modes embedded in structurally causal augmented data.