Early warning of nerve agent release in large indoor environments based on Encoder-Decoder Coupling Physics-Informed Neural Network

Accurately perceiving spatial distribution patterns, detecting dynamic evolution characteristics, and issuing early warnings set higher standards for indoor safety and protection efforts. Computational fluid dynamics (CFD) methods provide accurate predictions but struggle with real-time performance, while neural networks offer rapid predictions, though their performance declines with high-dimensional fluid flow. Nerve agents were used in a subway attack in Japan. For counter-terrorism surveillance purposes, the real-time and accurate detection of nerve agents is critical for providing early warnings to the public and coordinating subsequent rescue operations. Therefore, in this work, a new model called Encoder-Decoder Coupling Physics-Informed Neural Network is proposed, which learns from concentration data generated by experimentally validated CFD simulations and the spatio-temporal information to solve high-dimensional partial differential equations and provide predictions of nerve agents distribution that more closely align with objective physical laws. Extensive experiments are conducted, and the results indicate that our model attains the highest performance among all the algorithms proposed in this paper. The difference between the actual and predicted results is small, and the scatter points are distributed around the fitted curves. In addition, it can make predictions with millisecond-level response time to achieve real-time monitoring. We propose a data-physics dual-driven surrogate model for real-time monitoring and early warning of the distribution of nerve agents in indoor environments.