Manifold Learning with Implicit Physics Embedding for Reduced-Order Flow-Field Modeling

Nonlinear manifold learning (ML) based reduced-order models (ROMs) can substantially improve the quality of nonlinear flow-field modeling. However, noise and the lack of physical information often distort the dimensionality-reduction process, reducing the robustness and accuracy of flow-field prediction. To address this problem, we propose a novel manifold learning ROM with implicit physics embedding (IPE-ML). Starting from data-driven manifold coordinates, we incorporate physical parameters (e.g., angle of attack, Mach number) into manifold coordinates system by minimizing the prediction error of Gaussian process regression (GPR) model, thereby fine-tuning the manifold structure. These adjusted coordinates are then used to construct a flow-fields prediction model that predict nonlinear flow-field more accurately. The method is validated on two test cases: transonic flow-field modeling of the RAE2822 and supersonic flow-field modeling of the hexagon airfoil. The results indicate that the proposed IPE-ML can significantly improve the overall prediction accuracy of nonlinear flow fields. In transonic case, shock-related errors have been notably reduced, while in supersonic case the method can confine errors to small local regions. This study offers a new perspective on embedding physical information into nonlinear ROMs.

💡 Research Summary

This paper introduces a novel reduced‑order modeling (ROM) framework called Implicit Physics‑Embedding Manifold Learning (IPE‑ML) that integrates physical parameters directly into a nonlinear manifold representation to improve the prediction of complex flow fields. Traditional ROMs such as POD and DMD rely on linear subspaces and struggle with strongly nonlinear phenomena like shock waves and flow separation. Recent nonlinear approaches using autoencoders or manifold learning (e.g., Isomap, LLE) alleviate some issues but still treat physical parameters (angle of attack, Mach number, etc.) only as interpolation variables, leaving the learned manifold unaware of the underlying physics. Moreover, data‑driven manifold extraction can be corrupted by noise, degrading downstream predictions.

IPE‑ML addresses these shortcomings through a five‑step pipeline. First, a set of high‑fidelity CFD simulations is generated over a sampling space of physical parameters Θ (e.g., α, M∞). Second, an initial low‑dimensional coordinate matrix Y⁰ is obtained purely from data using either Isomap (global geodesic‑distance preserving) or Local Linear Embedding (local topology preserving). Third, a surrogate model f⁰ is built by training a Gaussian Process Regression (GPR) model that maps Θ → Y⁰. GPR is chosen for its Bayesian nature, providing smooth predictions and uncertainty estimates.

The core of IPE‑ML lies in step four: the manifold coordinates themselves are refined by minimizing the discrepancy between the GPR‑predicted coordinates Ŷ and the current coordinates Y. The loss function ½‖Y − Ŷ‖² is optimized with L‑BFGS, exploiting the analytically known gradient (Y − Ŷ). During each iteration the surrogate model is re‑trained, yielding an updated pair (*Y, *f) that jointly encode physical information. This implicit embedding forces the manifold to align with the physics‑driven surrogate, thereby correcting distortions caused by noise or purely data‑driven learning.

Finally, step five reconstructs full‑order flow fields for new parameter sets Θₚ. The refined GPR *f predicts the low‑dimensional coordinates Yₚ, and an inverse mapping based on kernel ridge regression (KRR) reconstructs the high‑dimensional field Xₚ. A Matérn kernel is employed in the KRR to provide greater flexibility than the standard RBF, while ridge regularization (λ) mitigates overfitting. The reconstruction follows X ≈ Φ(Y) = Z W Y, where Z is the kernel matrix between training and predicted coordinates, and W is obtained analytically via ridge regression.

The methodology is validated on two benchmark cases. The first is the transonic RAE2822 airfoil, where shock position and strength are highly sensitive to α and Mach number. IPE‑ML reduces shock‑related L₂ errors by more than 30 % compared with a conventional ML+GPR approach and lowers the overall mean error by roughly 20 %. The second case involves a supersonic hexagon airfoil, characterized by localized compression and expansion waves. Here IPE‑ML confines prediction errors to small regions and achieves an average error of 0.015, substantially better than the baseline. Both cases use a manifold dimension d≈8–10, with GPR hyper‑parameters (noise level, kernel length scale) and ridge regularization selected via cross‑validation.

Key advantages of IPE‑ML include: (1) direct incorporation of physical parameters into the manifold, enabling more faithful representation of nonlinear flow physics; (2) the use of GPR as a surrogate, which supplies uncertainty quantification; (3) a principled inverse mapping using kernel ridge regression with a Matérn kernel, offering flexible nonlinear reconstruction. Limitations are acknowledged: the additional GPR training and L‑BFGS optimization increase computational cost relative to purely data‑driven manifold learning; the performance is sensitive to the choice of manifold dimension and GPR kernel hyper‑parameters, suggesting a need for automated selection strategies; and the current formulation handles only static parameters, leaving time‑dependent or multi‑physics extensions for future work.

In conclusion, IPE‑ML demonstrates that embedding physics implicitly within a nonlinear dimensionality‑reduction framework can substantially enhance ROM accuracy for high‑speed, high‑Mach flows. Future research directions include automated dimension and hyper‑parameter tuning, extension to time‑varying and multi‑parameter settings, exploration of alternative surrogates (e.g., deep neural networks), and real‑time deployment on hardware‑constrained platforms.