Review of deep learning-based aerodynamic shape surrogate models and optimization for airfoils and blade profiles

Physics of Fluids, Journal Year: 2025, Volume and Issue: 37(4)

Published: April 1, 2025

In recent years, deep learning technology has developed rapidly and shown great potential in the optimization of complex systems. aerodynamic shape optimization, traditional computational fluid dynamics experimental methods are limited due to issues efficiency cost. contrast, surrogate models have gradually become a new alternative their advantages nonlinear modeling, efficient computation, flexible design. These offer novel approaches through such as data regression, automatic differentiation, operator learning. This paper presents comprehensive review latest research progress field based on models, focusing key technologies, application cases, future development trends. The article first elaborates importance context airfoil blade profile introducing background motivation. Then, it discusses technologies challenges faced optimization. Subsequently, introduces detail model, including data- physics-drisven neural networks, Physics-Informed Neural Networks Deep Operator Networks, practical cases these networks Finally, looks into pointing out Kolmogorov–Arnold improving model accuracy interpretability, well types summarizes development.

Language: Английский

Physics-informed neural networks with hybrid Kolmogorov-Arnold network and augmented Lagrangian function for solving partial differential equations

Scientific Reports, Journal Year: 2025, Volume and Issue: 15(1)

Published: March 27, 2025

Physics-informed neural networks (PINNs) have emerged as a fundamental approach within deep learning for the resolution of partial differential equations (PDEs). Nevertheless, conventional multilayer perceptrons (MLPs) are characterized by lack interpretability and encounter spectral bias problem, which diminishes their accuracy when used an approximation function diverse forms PINNs. Moreover, these methods susceptible to over-inflation penalty factors during optimization, potentially leading pathological optimization with imbalance between various constraints. In this study, we inspired Kolmogorov-Arnold network (KAN) address mathematical physics problems introduce hybrid encoder-decoder model tackle challenges, termed AL-PKAN. Specifically, proposed initially encodes interdependencies input sequences into high-dimensional latent space through gated recurrent unit (GRU) module. Subsequently, KAN module is employed disintegrate multivariate set trainable univariate activation functions, formulated linear combinations B-spline functions purpose spline interpolation estimated function. Furthermore, formulate augmented Lagrangian redefine loss model, incorporates initial boundary conditions multiplier terms, rendering multipliers learnable parameters that facilitate dynamic modulation balance among constraint terms. Ultimately, exhibits remarkable generalizability in series benchmark experiments, thereby highlighting promising capabilities application horizons

Language: Английский

Physics-Informed Neural Networks in Polymers: A Review

Polymers, Journal Year: 2025, Volume and Issue: 17(8), P. 1108 - 1108

Published: April 19, 2025

The modeling and simulation of polymer systems present unique challenges due to their intrinsic complexity multi-scale behavior. Traditional computational methods, while effective, often struggle balance accuracy with efficiency, especially when bridging the atomistic macroscopic scales. Recently, physics-informed neural networks (PINNs) have emerged as a promising tool that integrates data-driven learning governing physical laws system. This review discusses development application PINNs in context science. It summarizes recent advances, outlines key methodologies, analyzes benefits limitations using for property prediction, structural design, process optimization. Finally, it identifies current future research directions further leverage advanced modeling.

Language: Английский