Efficient neural network training method for unsteady flow field prediction based on data pool

In recent years, neural network technology has made significant progress in the field of unsteady flow field prediction, leading to the development of many innovative methods. However, deep learning-based unsteady flow field prediction techniques typically rely on autoregressive models, which inevitably face the issue of error accumulation. Existing solutions often suffer from challenges such as complex hyperparameter configurations and reduced training efficiency. To address these issues, this study makes the following contributions: (1) A novel training method is proposed to enhance model convergence by refining the training strategies employed. This approach achieves improved performance without necessitating complex hyperparameter configurations. (2) An innovative curriculum learning-based timestep reset strategy is introduced. This strategy further improves the convergence of neural networks and enhances prediction accuracy. A detailed comparative study was conducted on different training methods within the architecture of convolutional neural networks. Experimental results show that the proposed training strategy significantly improves prediction accuracy, with an improvement of up to an order of magnitude. Moreover, even when the training set spans only 600 timesteps, the model remains stable when predicting up to 9000 timesteps. Finally, our method also demonstrates high efficiency in terms of training time.