03.1 - Aerodynamics – CFD Methods and Validation
PHYSICS-INFORMED NEURAL OPERATORS FOR DISTURBANCE FIELD PREDICTION IN NONPARALLEL BOUNDARY-LAYER STABILITY ANALYSIS
G. Dalinajara Abreu Smith Borges¹, V. Silva Figueredo¹, G.L. Olichevis Halila¹, J. Messias Magalhães Júnior, Georgia Institute of Technology, United States; J.L. F. Azevedo, Instituto Tecnológico de Aeronáutica, Brazil; ¹Pontifícia Universidade Católica do Paraná, Brazil
Physics-informed data-driven modeling for flow stability analysis represents an inexpensive, easy-to-rnoperate, and easy-to-automate technique for complex flow phenomena. While standard neural net-rnworks are used as surrogates for computational fluid dynamics (CFD) simulations, they require largernvolumes of data, and the associated errors may not be acceptable. Physics-informed neural networksrn(PINNs) address this by using the underlying flow physics, represented by partial differential equa-rntions, to enhance modeling accuracy and enable successful training even from sparse data. This workrninvestigates the suitability of different PINN architectures for recovering disturbance fields in nonlocal,rnnonparallel flow stability analysis. The objective is to benchmark these architectures against data fromrna Parabolized Stability Equations (PSE) solver for a flat plate incompressible boundary layer flow. Thernanalysis will be conducted across a range of Reynolds numbers to assess the robustness, accuracy,rnand limitations of each model. As a proof of concept, preliminary results are presented which indicaternthat a standard PINN can accurately recover relevant stability metrics, including a precise reconstruc-rntion of the associated eigenfunctions. These results show that the PINN predictions accurately capturernthe disturbance’s amplitude spatial behavior, confirming the method’s viability even when only sparserntraining data is available.
DOI: 10.71945/icas2026_0051