A physics-informed neural network method for pressure-driven turbine-blade optimization

Yuan Huang; Francesca di Mare

Home/Archive/arXiv

arXiv · 2611.05691

A physics-informed neural network method for pressure-driven turbine-blade optimization

Yuan Huang·Francesca di Mare

View original source ↗Read source document ↗Export metadata ↓

Abstract

We develop a physics-informed method for aerodynamic shape optimisation of turbine blades, treating the flow field and blade geometry as coupled unknowns. The method translates prescribed surface-pressure distributions into geometric modifications through pressure-target continuation and alternating flow-geometry updates. A differentiable representation of the blade surfaces allows their positions and orientations to be optimised within the current flow field, followed by flow re-adaptation to the modified boundaries. The inverse formulation uses the two-dimensional steady compressible Euler equations as physics constraints and is evaluated on the LS89 turbine cascade using 28 pressure-target points derived from experimental data. Starting from a modified profile and a warm-started flow network, the final pressure-side and suction-side root-mean-square fitting errors are 1.13% and 2.21% of inlet total pressure, respectively. Independent forward verification with SharC reproduces the principal surface-pressure trends of the experimental measurements, with remaining differences near the trailing edge. These results demonstrate the translation of experimental-origin loading data into geometric blade modifications through coordinated optimisation of the flow field and moving boundaries.

Explore related subjects

physics.flu-dyn

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Yuan Huang, Francesca di Mare. 2026-11-03. A physics-informed neural network method for pressure-driven turbine-blade optimization. https://arxiv.org/abs/2611.05691

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Geometry-induced self-excited dynamo in a regular tetrahedron

We present a rotation-free magnetohydrodynamic dynamo driven by laminar thermal convection in a regular tetrahedral cavity. The tetrahedral boundaries organize the convective flow into a robust pattern of helical convection cells without global rotation or turbulence. Direct numerical simulations demonstrate exponential amplification of a weak seed magnetic field followed by nonlinear saturation. In the reference calculation, the saturated magnetic energy exceeds the kinetic energy. The velocity field develops $D_4$ dihedral symmetry, while the self-generated magnetic field exhibits a corresponding signed $D_4$ symmetry, including antisymmetry under $π$ rotations about the two horizontal axes. Analysis of the velocity and magnetic-field structures reveals a closed induction cycle sustained by geometry-induced helical convection. This system provides a conceptually simple setting for isolating and understanding the fundamental physical processes underlying magnetohydrodynamic dynamo action.

physics.flu-dyn↗

arXiv2026

Projection-based segregated, iterative, and multi-stage physics-informed neural network for incompressible Navier-stokes equations

Physics-informed neural networks(PINNs) offer a mesh-free and continuously differentiable approach to fluid simulation. However, standard monolithic PINNs for incompressible Navier-Stokes (NS) equations often suffer from severe constraint coupling and spectral biases, limiting both convergence and accuracy. To address this challenge, we propose a segregated, iterative, and multi-stage PINN (SIM-PINN) inspired by the classical projection method in computational fluid dynamics. In SIM-PINN, the nonlinear NS equations are decoupled into the linearized momentum and pressure Poisson equations. Each is addressed separately using dedicated multi-stage neural networks and coupled through projection-based inner iterations. Each stage of the multi-stage neural networks for the momentum equations corresponds to one iteration of the projection method, in which the velocity fields are progressively refined at targeted frequencies. Upon iteration convergence, solutions to the fully coupled and nonlinear NS equations are recovered. By integrating the segregated and iterative framework with the multi-stage training, SIM-PINN alleviates the optimization difficulties of the incompressible NS equations arising from nonlinearity, constraint coupling, and spectral biases. The improved convergence and accuracy are demonstrated by simulations of steady lid-driven cavity and unsteady Kolmogorov flows.

physics.flu-dyn↗

arXiv2026

Bayesian Kolmogorov-Arnold networks for trustworthy Reynolds-stress closures

Data-driven turbulence closures can improve Reynolds-averaged Navier-Stokes (RANS) predictions, but their use is limited by two issues: learned constitutive relations are often difficult to inspect, and predictions typically lack reliable uncertainty estimates. We develop a Bayesian tensor-basis Kolmogorov-Arnold network (BKAN) for Reynolds-stress anisotropy that addresses both within a single closure. Learnable univariate spline functions on network edges provide an inspectable representation that can, where sufficiently simple, be approximated symbolically. Variational posteriors over spline coefficients yield a predictive distribution separating epistemic uncertainty, due to limited training data, from aleatoric uncertainty, the irreducible scatter of high-fidelity targets at given inputs. Galilean and rotational invariance, symmetry, and tracelessness are enforced through the tensor basis, while physical realisability is treated separately through soft regularisation. A controlled regression benchmark assesses predictive accuracy, calibration, and Bayesian parameterisation. Placing the posterior on spline coefficients gives the best overall balance of accuracy, calibration, parameter cost, and interpretability. Applied to Reynolds-stress anisotropy, the BKAN is accurate in distribution, with calibrated, informative uncertainty that tends to identify larger prediction errors. Under geometric extrapolation, however, the model becomes overconfident: epistemic uncertainty increases but does not fully account for growing prediction errors. Finally, the framework is extended to three-dimensional wind-farm wakes, where the learned correction transfers to unseen layouts and posterior samples are propagated through the RANS solver, with the aleatoric component injected as a spatially correlated field, to obtain uncertainty bounds on the predicted wake flow.

physics.flu-dyn↗