Scholarlink

<p>Rapid acquisition of high-fidelity flow fields and performance metrics using neural network models is essential for flow analysis and engineering applications. This study develops a hybrid model, termed U-LSTM, which integrates a U-shaped neural network (U-NET) with a long short-term memory (LSTM) network to predict high-dimensional, unsteady flow fields. The model accurately captures both flow evolution and force coefficients of flapping wings. To enhance training efficiency, an improved bayesian optimization and hyperband (BOHB) algorithm is incorporated for automated hyperparameter tuning. Results show that U-LSTM accelerates the training process and achieves mean absolute errors (MAE) below 5 % in predicting flow fields across different phases of the flapping cycle. The integration of BOHB further reduces hyperparameter tuning and training costs while improving model stability and robustness. Moreover, a reduced-scale version of the U-LSTM (RUL) achieves absolute errors below 0.25 % in force coefficient prediction, with lower computational demand compared to other models, demonstrating the strong generalization capability of the U-LSTM architecture.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Neural network model based on Bayesian optimization and hyperband algorithm for predicting flapping wing’s flow field and force coefficients

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Jinlong Gu,
  • Feilin Wang,
  • Di Jiang,
  • Long Chang,
  • Bing Zhu

摘要

Rapid acquisition of high-fidelity flow fields and performance metrics using neural network models is essential for flow analysis and engineering applications. This study develops a hybrid model, termed U-LSTM, which integrates a U-shaped neural network (U-NET) with a long short-term memory (LSTM) network to predict high-dimensional, unsteady flow fields. The model accurately captures both flow evolution and force coefficients of flapping wings. To enhance training efficiency, an improved bayesian optimization and hyperband (BOHB) algorithm is incorporated for automated hyperparameter tuning. Results show that U-LSTM accelerates the training process and achieves mean absolute errors (MAE) below 5 % in predicting flow fields across different phases of the flapping cycle. The integration of BOHB further reduces hyperparameter tuning and training costs while improving model stability and robustness. Moreover, a reduced-scale version of the U-LSTM (RUL) achieves absolute errors below 0.25 % in force coefficient prediction, with lower computational demand compared to other models, demonstrating the strong generalization capability of the U-LSTM architecture.