The accurate deformation of computational meshes plays a crucial role in CFD simulations, ensuring numerical stability and preserving grid quality without resorting to inefficient and costly remeshing procedures. With the rapid development and growing accessibility of machine learning, neural networks are increasingly being integrated into CFD workflows. This thesis presents a novel physics-informed neural network–based approach for mesh deformation. The proposed framework employs two neural networks: the first predicts the global deformation field, while the second applies a corrective adjustment, followed by the exact enforcement of boundary conditions. Additionally, a boundary-layer preservation scheme is introduced to maintain near-wall resolution. The method demonstrates strong flexibility and robustness across both two- and three-dimensional cases, effectively handling large deformation magnitudes and configurations involving multiple moving objects while maintaining high mesh quality. A scalability analysis highlights the model’s efficiency in handling large datasets through GPU acceleration, while transfer learning and cross-resolution generalization further reduce computational costs. Comparative tests with ELA and RBF methods show that the proposed approach achieves comparable or superior mesh quality. An unsteady CFD simulation involving a complex geometry is also performed, further validating the effectiveness and robustness of the proposed deformation approach under realistic flow conditions. Although classical methods remain faster for small grids, the neural approach offers competitive performance and strong potential for large-scale CFD applications, outlining a promising research direction for the future development of adaptive and physics-informed mesh manipulation strategies in CFD.
La deformazione di mesh riveste un ruolo fondamentale nelle simulazioni CFD, garantendo la stabilità numerica e preservando la qualità della griglia senza ricorrere a procedure di remeshing inefficienti e costose. Con il rapido sviluppo e la crescente accessibilità delle tecniche di machine learning, le reti neurali stanno assumendo un ruolo sempre più importante nell’ambito della CFD. In questa tesi viene presentato un nuovo approccio basato sulle Physics-Informed Neural Networks (PINNs) per la deformazione delle mesh. Il metodo proposto si fonda su un’architettura a doppia rete neurale, in cui una prima rete apprende il campo di deformazione globale mentre la seconda applica uno spostamento correttivo; in seguito vengono imposte le condizioni al contorno in modo esatto. È inoltre introdotto un meccanismo di conservazione della risoluzione della mesh nelle regioni a parete, fondamentali per il corretto calcolo e sviluppo dello strato limite. Il metodo dimostra grande flessibilità e robustezza sia in due che in tre dimensioni, mantenendo una qualità di griglia comparabile o superiore anche in presenza di ampie deformazioni e configurazioni con più superfici in movimento. L’analisi di scalabilità evidenzia l’efficienza del modello nell’elaborazione di dataset di grandi dimensioni grazie all’impiego della GPU, mentre tecniche di transfer learning e di generalizzazione tra griglie di diversa risoluzione consentono di ridurre ulteriormente i tempi computazionali. Una simulazione CFD instazionaria condotta su una geometria complessa conferma ulteriormente l’efficacia dell’approccio proposto in condizioni realistiche. Nonostante le tecniche classiche esibiscano tempi computazionali inferiori, i risultati ottenuti confermano le potenzialità di questo approccio come promettente direzione di ricerca futura nell'ambito della deformazione di griglia.
Physics-informed neural networks for mesh deformation
Marinaro, Giorgio
2024/2025
Abstract
The accurate deformation of computational meshes plays a crucial role in CFD simulations, ensuring numerical stability and preserving grid quality without resorting to inefficient and costly remeshing procedures. With the rapid development and growing accessibility of machine learning, neural networks are increasingly being integrated into CFD workflows. This thesis presents a novel physics-informed neural network–based approach for mesh deformation. The proposed framework employs two neural networks: the first predicts the global deformation field, while the second applies a corrective adjustment, followed by the exact enforcement of boundary conditions. Additionally, a boundary-layer preservation scheme is introduced to maintain near-wall resolution. The method demonstrates strong flexibility and robustness across both two- and three-dimensional cases, effectively handling large deformation magnitudes and configurations involving multiple moving objects while maintaining high mesh quality. A scalability analysis highlights the model’s efficiency in handling large datasets through GPU acceleration, while transfer learning and cross-resolution generalization further reduce computational costs. Comparative tests with ELA and RBF methods show that the proposed approach achieves comparable or superior mesh quality. An unsteady CFD simulation involving a complex geometry is also performed, further validating the effectiveness and robustness of the proposed deformation approach under realistic flow conditions. Although classical methods remain faster for small grids, the neural approach offers competitive performance and strong potential for large-scale CFD applications, outlining a promising research direction for the future development of adaptive and physics-informed mesh manipulation strategies in CFD.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_Marinaro_Thesis_01.pdf
accessibile in internet per tutti
Descrizione: testo tesi
Dimensione
106.78 MB
Formato
Adobe PDF
|
106.78 MB | Adobe PDF | Visualizza/Apri |
|
2025_12_Marinaro_Executive_Summary_02.pdf
accessibile in internet per tutti
Descrizione: executive summary
Dimensione
2.52 MB
Formato
Adobe PDF
|
2.52 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/247125