Physics-constrained neural networks for rapid performance evaluation of laser atmospheric propagation

Laser atmospheric propagation efficiency assessment is a key technology in laser engineering applications, the core of which is to quickly and accurately assess the transmission effect of laser beams in complex atmospheric environments. Aiming at the problem of low computational efficiency of fluctuating optical model and insufficient integration of physical mechanism of machine learning model in the traditional laser atmospheric propagation evaluation method, this paper proposes a neural network model that integrates physical laws (PINN model). The model introduces physical constraints such as monotonicity constraints, negative correlation constraints, and self-consistency constraints in laser atmospheric propagation, and embeds the physical a priori knowledge into the neural network framework to avoid the black-box-type limitations that exist in the purely data-driven model, and at the same time, improves the physical reasonableness of the model's prediction results. The experimental results show that the prediction error of the PINN model is 40.4% lower than that of the traditional data-driven model, and the average inference time of single data is kept within 5 μs, which meets the real-time evaluation requirements in engineering applications. The research results can provide technical support and solutions for the design and application of optoelectronic engineering systems.