Cascaded LSTM with downsampling strategy for fast and accurate modeling of chirped pulse amplification

Conventional iterative numerical modeling of chirped pulse amplification (CPA) systems requires a large temporal simulation window to accommodate the heavily chirped pulses. Simultaneously, maintaining the high temporal resolution certainly improves the computational complexity, posing challenges for CPA system design and optimization based on numerical simulations. To overcome this limitation, we propose a cascaded long-short-term memory (LSTM) model with a downsampling strategy trained for efficient and accurate modeling of a multi-stage optical fiber system. This approach delivers full-field simulation of the heavily chirped pulse with 10-nm spectral bandwidth at the pulse energy reaching 14.9 μJ. Through aggressive downsampling in the time domain, the proposed framework reduces the computational complexity by 929 times and achieves a remarkable 1,564-fold speedup compared to conventional numerical simulations, while maintaining prediction errors of the pulse energy and duration below 2%. Our work provides an efficient and high-fidelity CPA systems modeling alternative, which is particularly suitable for the inverse design and optimization of CPA systems for high-energy short pulses generation.