Scholarlink

<p>This study compares the predictive performance of standard Neural Networks (NNs) and Physics-Informed Neural Networks (PINNs) in estimating pressure drops across laminar and turbulent flow regimes. The proposed PINN framework integrates the Hagen–Poiseuille and Darcy–Weisbach equations, incorporating multiple friction factor correlations (Blasius, Swamee–Jain, and Haaland) to evaluate the impact of physical constraints and regularization on model generalization. The dataset, generated from established analytical models, covers a wide range of industrially relevant parameters (pipe diameter 0.005–0.1&#xa0;m, velocity 0.5–20&#xa0;m/s). Results show that PINNs substantially outperform standard NNs, achieving up to 41% lower Mean Absolute Error (MAE) while maintaining strong generalization across unseen data. The optimal PINN configuration (α = 0.0001, Blasius model) achieved an MAE of 1.31, compared to 2.23 for the best NN optimized by Bayesian search. These findings demonstrate that embedding physical laws enhances predictive accuracy and interpretability, especially under limited data conditions. The developed framework offers a robust and computationally efficient tool for mechanical system design, pipeline optimization, and fluid transport analysis.</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

Physics-informed neural network framework for accurate pressure drop prediction in laminar and turbulent pipe flows

  • 期刊论文

  • 发表日期：

    2026年5月16日

查看全文
  • Iwan H. Sahputra,
  • Hariyo P. S. Pratomo,
  • Ekadewi A. Handoyo,
  • Indar Sugiarto,
  • Egawati Panjei

摘要

This study compares the predictive performance of standard Neural Networks (NNs) and Physics-Informed Neural Networks (PINNs) in estimating pressure drops across laminar and turbulent flow regimes. The proposed PINN framework integrates the Hagen–Poiseuille and Darcy–Weisbach equations, incorporating multiple friction factor correlations (Blasius, Swamee–Jain, and Haaland) to evaluate the impact of physical constraints and regularization on model generalization. The dataset, generated from established analytical models, covers a wide range of industrially relevant parameters (pipe diameter 0.005–0.1 m, velocity 0.5–20 m/s). Results show that PINNs substantially outperform standard NNs, achieving up to 41% lower Mean Absolute Error (MAE) while maintaining strong generalization across unseen data. The optimal PINN configuration (α = 0.0001, Blasius model) achieved an MAE of 1.31, compared to 2.23 for the best NN optimized by Bayesian search. These findings demonstrate that embedding physical laws enhances predictive accuracy and interpretability, especially under limited data conditions. The developed framework offers a robust and computationally efficient tool for mechanical system design, pipeline optimization, and fluid transport analysis.