Real-time physics-informed reconstruction of transient fields using sensor guidance and higher-order time differentiation

Abstract

This study proposes FTI-PBSM (Fixed-Time-Increment Physics-informed neural network-Based Surrogate Model), a novel physics-informed surrogate modeling framework designed for real-time reconstruction of transient responses in time-dependent Partial Differential Equations (PDEs) using only sparse, time-dependent sensor measurements. Unlike conventional Physics-Informed Neural Network (PINN)-based models that rely on Automatic Differentiation (AD) over both spatial and temporal domains and require dedicated causal network architectures to impose temporal causality, the proposed approach entirely removes AD in the time direction. Instead, it leverages higher-order numerical differentiation methods, such as the Central Difference, Adams-Bashforth, and Backward Differentiation Formula, to explicitly impose temporal causality. This leads to a simplified model architecture with improved training stability, computational efficiency, and extrapolation capability. Furthermore, FTI-PBSM is trained on sparse sensor measurements from multiple PDE cases generated by varying PDE coefficients, with the sensor data serving as model input. This enables the model to learn a parametric PDE family and generalize to unseen physical cases, accurately reconstructing full-field transient solutions in real time. The proposed model is validated on four representative PDE problems-the convection equation, diffusion-reaction dynamics, Korteweg-de Vries (KdV) equation, and Allen-Cahn equation-and demonstrates superior prediction accuracy and generalization performance compared to a causal PBSM, which is used as the baseline model, in both interpolation and extrapolation tasks. It also shows strong robustness to sensor noise and variations in training data size, while significantly reducing training time.