Long-range prediction of temperature-tunable soliton dynamics using WaveNet-LSTM based on feature fusion

Precise prediction and on-demand control of soliton state switching in temperature-tunable ultrafast lasers are pivotal for the development of next-generation intelligent photonic systems. However, this task is hindered by the prohibitive computational costs of traditional numerical simulations. Concurrently, existing deep learning models often fail to capture thermally induced transient dynamics and lack the capability to accurately model long-range evolution. To address these challenges, a feature fusion-based WaveNet-LSTM (WN-LSTM) model is proposed. By integrating external temperature parameters with optical pulse complex-field data, the model directly learns the physical mapping between thermal regulation and soliton evolution. Leveraging a unique parallel processing architecture, it efficiently captures both the transient switching events and the long-range evolutionary dynamics. The model demonstrates the successful prediction of complex soliton state transitions triggered by temperature variations in both normal and anomalous dispersion regimes. The predictions exhibit high consistency with numerical simulations and are rigorously validated experimentally using a real temperature-tunable fiber laser. A comparative analysis of NRMSE and training time reveals that the WN-LSTM model significantly outperforms traditional LSTM and CRNN architectures in both accuracy and efficiency. Consequently, this work presents a powerful, data-driven tool for the intelligent design and on-demand control of ultrafast lasers.