PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
February 22, 2026Japanese Journal of JSCE0 citations

Application of Physics-Informed Neural Networks to Flows in Non-Straight Channel Geometries

View Full Paper
YOYuki OHARASSShohei SEKIDMDaichi MOTEKI

Key Points

  • The aim is to investigate the applicability of physics-informed neural networks (PINNs) for modeling flows in non-straight channel geometries.
  • Applied PINNs to flows in converging and meandering channels.
  • Compared PINNs estimations with numerical computations under the same conditions.
  • Used ensemble models of PINNs to check consistency of estimations.
  • Observed good agreement in depth and velocity distributions for non-linear channels.
  • Identified underestimation of flow velocities in converging channels due to spectral bias.
  • Noted the requirement of high frequency component learning and loss design for accurate modeling.

Abstract

複雑な平面形状の河川では,数値計算やデータ同化で一般座標が頻用されるが,格子由来の誤差が問題となる.本研究では,格子生成が不要なPhysics-informed neural networks(PINNs)を漸縮水路と蛇行水路の流れに適用し,妥当性と課題を検討した.PINNsの推定と同一条件の数値計算を比較した結果,非直線水路で水深・流速分布の良好な一致を確認した.漸縮水路ではスペクトルバイアスにより流速が過小推定され,連続式を満たすため水深が過大となる誤差が生じた.蛇行水路の砂州の上流端では,境界条件の流量フラックスを満たすため流速が過大となった.また,複数のPINNsモデルをアンサンブル的に用いることで再現性のある推定を確認した.PINNsは非直線形状水路の流れに有効だが,高周波数成分の学習や損失設計の重要性が示唆された.

Ask AI
Helpful
Bookmark
Share
View Full Paper

Cite This Study

OHARA et al. (2026) studied this question.

synapsesocial.com/papers/699a9ceb482488d673cd296fhttps://doi.org/10.2208/jscejj.25-16093
Ask AI
Helpful
Bookmark
Share
View Full Paper