Spatiotemporal Sequence Forecasting of Tin Droplet-Based Laser Produced Plasma Physical Fields Using Fourier Neural Operator

Deep learning techniques play a pivotal role in predicting the spatiotemporal evolution of physical fields in tin droplet-based laser-produced plasma (LPP), which is essential for advancing our understanding and control of plasma behavior in extreme ultraviolet lithography (EUVL) sources. To accelerate the modeling of such complex systems at a significantly lower computational cost compared to traditional numerical methods, we present the first application of neural operators to twodimensional compressible magnetohydrodynamics (MHD) simulations for LPP-based EUVL light sources. In this study, extensive two-dimensional radiation hydrodynamic simulations were performed to generate raw plasma physical field data for LPPs. Our primary objective is to leverage the Fourier Neural Operator (FNO) for accurate and efficient forecasting of the spatiotemporal evolution of the plasma, while also conducting comparative experiments with Convolutional Long Short-Term Memory (ConvLSTM) and Convolutional Neural Operator (CNO) networks. This iterative approach informs the design and control of future laser-plasma EUVL source experiments. The FNO model is particularly well-suited for solving multivariate partial differential equations, as it captures complex variable interdependencies within the system.