Neural Operator-Based Surrogate Modeling for Real-Time Multiphysics Simulation and Control

Abstract

Multiphysics simulations are widely used in the modelling of engineering and scientific systems with complex geometries; where traditional numerical methods like finite element and finite difference methods have a high computational cost, their application in real-time is very limited. In this paper a neural operator-based surrogate modeling framework will be introduced that allows to perform efficient and accurate real-time multiphysics simulation and control. The proposed method takes advantage of state-of-the-art operator learning methods, such as the Fourier Neural Operator (FNO), Deep Operator Network (DeepONet), and Physics-Informed Neural Operator (PINO) to train the maps between input parameter spaces and the solution fields of governing partial differential equations. The framework combines both a high predictive accuracy and high physical consistency by incorporating the use of data-driven learning and physical constraints. Massive testing of the models has shown that the models are able to attain very high levels of computational latency reduction, with high fidelity, measured by indices like Relative L2 Error, Root Mean Squared Error (RMSE). Moreover, the surrogate model is smoothly combined with a control module, which allows making decisions in real time and optimising the system. The findings demonstrate that neural operator-based surrogates can be an effective solution to close the computational efficiency to physical accuracy gap, making them a potential solution to the next-generation real-time Multiphysics simulator and control system.