Source-linked AI summary
A Residual Learning Approach for Unsteady Aerodynamic Load Prediction
Divya Sanghi, Carlos E. S. Cesnik
TL;DR
Unsteady aerodynamic load prediction needs models that capture effects beyond classical linear, incompressible baselines. This paper combines a Wagner-based residual LSTM with CFD data and finds lower average external-test error and stronger withheld-family generalization than direct LSTM prediction, although high-frequency pitch-only cases favor the direct model.
Problem
Whether an analytical low-fidelity model can provide an effective residual baseline across motion types and distribution shifts remains insufficiently studied.
Method
The study trains an LSTM to predict the difference between CFD lift and a static-corrected Wagner baseline, comparing it with a direct CFD-lift LSTM.
Results
0.054 ± 0.007 versus 0.065±0.009 NRMSE: the residual model has lower mean external-benchmark error and generally stronger leave-family-out generalization than the direct model.
Takeaways & Limitations
Residual learning shows promise as a modular augmentation of classical low-order aerodynamic theories when the baseline removes a structured response component.
Takeaways & Limitations
The direct model remains better for the high-frequency pitch-only case, where the Wagner baseline does not simplify the residual target enough.
Abstract
from arXiv · showhide
This paper investigates the feasibility of using residual learning to improve unsteady aerodynamic load prediction for aeroelastic applications. The machine learning technique selected for the study is the long short-term memory (LSTM) neural network, which is used for its suitability for sequential data with aerodynamic memory effects. The approach is investigated for the NLR 7301 airfoil benchmark using high-fidelity CFD lift data for prescribed pitch and plunge motions in the transonic flow regime in the presence of shock motion. An analytical unsteady aerodynamic model based on the Wagner function is used as a physics-based baseline, and the neural network is trained to learn the difference between the CFD lift coefficient and the Wagner prediction. The residual model is compared with a direct neural-network model trained to predict the CFD lift coefficient. The comparison includes feature and normalization studies, external benchmark cases, and leave-one-out and leave-family-out generalization tests across a range of sinusoidal and non-sinusoidal motions. The residual model performs best when its inputs align with the Wagner formulation variables, generally giving lower error and more consistent performance across training runs, though the direct model remains more accurate for some high-frequency cases. The residual model also generalizes better in the leave-one-out and leave-family-out tests, with a smaller increase in error than the direct model when entire motion families are withheld from training. Overall, the results indicate that residual learning shows promise as a modular approach for augmenting classical low-order aerodynamic theories, especially when the physics baseline removes a structured part of the aerodynamic response and leaves a lower-variance correction for the neural network to learn.
1 Introduction
The introduction motivates physics-guided residual learning for sequential unsteady aerodynamic-load prediction and identifies an open question: whether an analytical low-fidelity model can provide an effective residual baseline across diverse motions. It positions the Wagner function as a practical baseline whose linear, incompressible, attached-flow assumptions limit applicability.
- Motivation: Unsteady aerodynamic loads depend on instantaneous kinematics, reduced frequency, and prior motion history, making accurate prediction important for aeroelastic analysis, flutter assessment, and control-oriented modeling.Classical theories such as Theodorsen’s formulation and the Wagner function separate attached-flow responses into circulatory and noncirculatory components.
- Data-driven modeling: Because unsteady lift is sequential, recurrent neural networks can represent dependence on current motion variables and prior motion history while avoiding repeated governing-equation solves during online prediction.Neural-network reduced-order models learn aerodynamic input–output behavior directly from data.
- Benchmark and prior work: The NLR 7301 airfoil is an established benchmark for data-driven unsteady aerodynamic modeling, including neural-network predictions of pitch–plunge loads, lift histories, and pressure distributions.Prior studies explored radial-basis-function, improved-ROM, and multivariate recurrent approaches for this benchmark.
- Research gap: Whether an analytical low-fidelity model can serve as an effective residual baseline across a broad range of motion types remains insufficiently studied.The Wagner function requires prescribed airfoil kinematics and flow speed without fitting data or CFD solves, but its linear, incompressible, and attached-flow assumptions constrain applicability.
2 Benchmark Problem and Physics Baseline
The benchmark uses high-fidelity CFD for transonic pitching and plunging motions of the NLR 7301 airfoil, with Wagner-based analytical models providing the physics baseline. Linear unsteady Wagner is selected after comparison with quasi-steady and nonlinear variants, with a static correction applied before residual calculation.
- Benchmark case: The NLR 7301 benchmark uses CFD data for pitching and plunging motions at Reynolds number 2.1 × 10^6 and Mach numbers 0.65–0.75, including moving transonic shocks.The dataset was generated with the in-house xflow CFD code.
- Dataset and benchmarks: The training dataset contains 48 cases across four prescribed-motion families, including single-harmonic, amplitude-modulated, Gaussian-pulse, and noise-perturbed harmonic inputs.The families cover controlled frequency and Mach variation as well as non-sinusoidal and transient behavior.
- Dataset and benchmarks: Three external benchmark cases are retained for testing: high-frequency pure pitch, moderate-frequency pure plunge, and low-frequency coupled pitch–plunge motion.These cases are not used for fitting and remain complete time histories in the same nondimensional time convention as the training data.
- Physics baseline: The Wagner baseline predicts unsteady lift from prescribed pitch–plunge kinematics and free-stream conditions using noncirculatory added-mass and circulatory contributions.Three variants are compared: quasi-steady, linear unsteady, and nonlinear unsteady Wagner.
- Physics baseline: A scalar static correction removes the steady lift offset from the incompressible Wagner prediction before defining the time-dependent residual relative to CFD lift.The raw baseline does not capture offsets caused by finite Mach number, airfoil shape, and nonzero mean angle of attack.
- Physics baseline: Linear unsteady Wagner is selected because quasi-steady predictions have the largest errors, while linear and nonlinear unsteady predictions are nearly identical on the external benchmarks.The comparison indicates that linear unsteady response dominates and nonlinear amplitude correction has no significant effect for these cases.
3 Residual Learning Methodology
The methodology decomposes lift prediction into a linear-Wagner physics baseline plus an LSTM residual correction and compares it with a direct LSTM model. It specifies distinct feature and normalization choices, case-based evaluation, and error metrics for both approaches.
- Residual-learning formulation: The residual framework represents aerodynamic output as a physics-based baseline plus learned residual modules, but this study evaluates one LSTM correction with fixed g1 = 1 and the linear Wagner model as baseline.The implemented case is a single-module unsteady-lift application; broader multi-module extensions are deferred.
- Residual-learning formulation: The residual LSTM learns the difference between CFD lift and the Wagner prediction, then adds its predicted correction to the Wagner baseline.This defines the residual target and the final residual-model lift prediction.
- Model implementation: Both direct and residual models use one LSTM layer with sequence output, a fully connected layer, and a regression output layer, trained and evaluated on original τ-indexed solver samples.Data splits are made by complete motion case rather than by individual time sample.
- Evaluation procedure: Lift-history errors are evaluated with RMSE and NRMSE, with the latter normalized by the CFD lift range over the same evaluation window.Unless otherwise stated, the evaluation window is the final 60% of each history to reduce startup-transient influence.
- Feature and normalization study: The selected direct model uses the compact rate-based input set [α, ˙α, ˙h, M]T with normalized output, whereas the residual model benefits from variables aligned with the Wagner function and input-only normalization.The feature and normalization study reflects the different prediction targets learned by the two models.
- Feature and normalization study: Adding second-derivative features raises the direct model’s mean NRMSE from 0.054 to 0.104, indicating that acceleration inputs can degrade direct-model performance.The passage attributes this sensitivity to noisy acceleration inputs.
4 Results
The residual model outperforms the direct model across most external-benchmark and generalization tests, while the direct model remains preferable for some pitch-only or low-frequency cases. Its advantage is strongest when the Wagner baseline supplies transferable response structure, although performance can be seed-sensitive for withheld transient families.
- External benchmark: 0.054 ± 0.007 versus 0.065±0.009: the residual model achieves lower mean external-benchmark NRMSE than the direct model across three unseen cases.The comparison averages results over five random seeds.
- External benchmark: The direct model remains better for the external pure-pitch case because the Wagner baseline does not sufficiently reduce phase and amplitude mismatch.The residual model shows smaller random-seed scatter on plunge-only and coupled cases.
- Leave-one-out generalization: The residual model gives lower NRMSE for most single-harmonic leave-one-out holdouts, with largest gains mostly at higher reduced frequency and Mach number.Nearby training cases allow the LSTM to refine phase structure during in-family interpolation.
- Leave-one-out generalization: The residual model has a lower family-mean NRMSE for all four leave-one-out motion families, despite fewer residual-favorable holdouts in the amplitude-modulated family.The result indicates that the residual advantage extends beyond the sinusoidal training family.
- Leave-family-out generalization: In leave-family-out testing, the residual-model mean remains close to its leave-one-out value when single-harmonic cases are withheld, whereas withholding that family mainly affects the direct model.The Wagner baseline supplies transferable sinusoidal response structure, while the direct model must learn it from non-sinusoidal examples.
- Leave-family-out generalization: The Gaussian-pulse leave-family-out case is most sensitive to random-seed choice, although the representative run favors the residual model.The residual advantage can weaken for transient cases that differ more strongly from the remaining training families.
5 Concluding Remarks
The study supports residual LSTM learning as a useful augmentation to classical unsteady aerodynamic models, particularly when the physics baseline removes structured aerodynamic response. The residual model generally improves generalization, although the direct model remains better for the high-frequency pitch-only case.
- Study scope: The study compared a direct LSTM predicting total CFD lift with a residual LSTM predicting the difference from a mean-aligned Wagner baseline.The evaluation used external benchmark tests, leave-one-out studies, and leave-family-out tests on the NLR 7301 airfoil in transonic flow with shock motion.
- Model and feature choices: The linear Wagner model performed nearly the same as the nonlinear variant and better than the quasi-steady baseline for external benchmark cases.The residual model favored derivative and acceleration features with input-only normalization, whereas the direct model favored a compact rate-based feature set.
- Limitation: The direct model remains better for the high-frequency pitch-only case because the Wagner baseline does not simplify the residual target enough to improve reconstructed lift prediction.This identifies a condition under which residual learning does not provide the best final prediction.
- Generalization: The residual model improves family-mean NRMSE for all four motion families and gives lower NRMSE for most holdouts in three leave-one-out families.The reported families are single-harmonic, Gaussian-pulse, and noise-perturbed motions.
- Implications and future work: Residual learning is most effective when the physics baseline removes structured aerodynamic response and leaves the neural network a smaller unsteady correction to learn.The paper proposes further study of Mach-number interpolation and extrapolation, replacement of the scalar static correction with a static ROM, and evaluation of alternatives.
Copyright statement
The authors confirm ownership of the paper’s original material and permissions for included third-party content and the paper’s publication and public distribution.
- The authors confirm that they or their organization hold copyright on all original material in the paper.
- The authors confirm obtaining permission from copyright holders for third-party material included in the paper.
- The authors confirm permission for publication and public distribution of the paper.