High-resolution multiphysics predictions and multifields reconstruction for chemical lasers enabled by operator neural networks

Abstract

Numerical simulation is an important approach for chemical laser investigation. However, it is time-consuming and prone to numerical divergence because of the system's complexity and strong nonlinearity. This paper develops a deep learning encoding underlying physics method for modeling chemical oxygen-iodine lasers (COIL). Two classes of DeepONets are proposed to capture the nonlinear relationships between flow and optical fields. The DeepONet’s capability of generalization outside the function space of the dataset is obtained, despite being trained by a noisy dataset. The high-frequency oscillations in the laser data increase the DeepONet prediction errors on optical quantities. Two solutions, subdomain division with sampling refinement and filtering high-frequency component of the laser by Fourier transform, are proposed to address this issue and achieve a nearly 50 % error reduction. The pre-trained DeepONet modules, functioning as building blocks, are utilized to construct parallel and serial DeepM&Mnets through which only 10 discrete data points among 22,877 are used to reconstruct all physical fields for COIL with errors below 10 %. The serial DeepM&Mnet method leverages limited laser measurement data to reconstruct flow field and component distributions, marking a significant step forward in solving the inverse problem in chemical laser systems.