Aug 2026· Proceedings of the Institution of Mechanical Engineers. Part G, Journal of Aerospace Engineering· 0 citations· 21 references
TL;DR
This work proposes an efficient multilayer neural network framework to predict the pitching moment coefficient of canard-controlled missiles, significantly reducing the need for costly CFD data and providing a powerful surrogate tool for rapid design optimization based on trim angle of attack and geometric parameters.
Abstract
Aerodynamic design of aerospace vehicles often necessitates extensive Computational Fluid Dynamics (CFD) simulations, which are computationally expensive. To address this, we propose an efficient multilayer neural network framework to predict the pitching moment coefficient of canard-controlled missiles. A low-fidelity database of over 28,000 points was rapidly generated using Missile DATCOM and used to train an initial neural network with four hidden layers. The core of our methodology is an architectural transfer learning approach, where this pre-trained model initializes a high-fidelity network, significantly reducing the need for costly CFD data (using only 120 samples). The Levenberg-Marquardt algorithm’s hyperparameters were fine-tuned to optimize performance. The final model achieved a Root Mean Square Error (RMSE) of 0.0055 on a random test dataset. The model’s stability and generalization capability were further confirmed through 5-fold cross-validation, which demonstrated robust performance. This highly accurate and validated model provides a powerful surrogate tool for rapid design optimization based on trim angle of attack and geometric parameters.
Aero-propulsive interactions in flapped configurations are a critical consideration for Short Take-Off and Landing (stol) aircraft, where extreme operational requirements demand robust, optimization-ready methodologies during preliminary design. This study develops a surrogate modeling framework that predicts the section-level aerodynamic response of a propeller–airfoil–flap configuration across a multi-dimensional space of propeller positioning, flap geometry, and operational conditions. A paired powered and unpowered design of experiments isolates the propulsive contribution to lift, drag, and pitching moment, while a virtual-disk propeller model parameterized by volumetric thrust decouples the prediction from any specific blade design. The framework couples two-dimensional steady Reynolds-averaged Navier–Stokes (rans) dataset generation with a Deep Neural Network (dnn) surrogate, which achieves coefficient of determination values above 0.96 for all three coefficients and reduces evaluation cost by several orders of magnitude relative to direct cfd, a benefit that is decisive in optimization. Single- and multi-objective optimization identify Pareto-optimal configurations, and independent cfd verification confirms prediction accuracies within 5 to 10% across the operational envelope. The resulting surrogate enables rapid, optimization-ready exploration of propeller–airfoil–flap configurations, providing actionable trade-off information for the preliminary design of stol aircraft.
Gabriele Morra, S. Corcione, F. Nicolosi· Applied Sciences· 0 citations
Accurate machine-learning models for aerodynamic prediction are essential for accelerating shape optimization yet remain challenging to develop for complex three-dimensional configurations due to the high cost of generating training data. This work introduces a methodology for efficiently constructing accurate surrogate models for design purposes by first pretraining a large-scale model on diverse geometries and then fine-tuning it with a few more detailed task-specific samples. A Transformer-based architecture, AeroTransformer, is developed and tailored for large-scale training to learn aerodynamics. The methodology is evaluated on transonic wings, where the model is pretrained on SuperWing, a dataset of nearly 30,000 samples with broad geometric diversity, and subsequently fine-tuned to handle specific wing shapes perturbed from the Common Research Model. Results show that, with 450 task-specific samples, the proposed methodology achieves a 0.36% error on surface-flow prediction, reducing error by 84.2% compared to training from scratch. The influence of model configurations and training strategies is also systematically studied to provide guidance on effectively training and deploying such models under limited data and computational budgets. To facilitate reuse, we release the datasets and the pretrained models at https://github.com/tum-pbs/AeroTransformer . An interactive design tool is also built on the pretrained model and is available online at https://webwing.pbs.cit.tum.de .
This paper introduces a framework for airfoil inverse design, using a deep-learning approach and incorporating both the parametric section (PARSEC) and class–shape transformation (CST) airfoil parameterization techniques for comparative evaluation. A predictive model has been established to estimate the PARSEC and CST parameters that characterize airfoil shapes based on a range of aerodynamic coefficients as inputs. This model is constructed through the training of a neural network using an aerodynamic dataset generated from XFOIL simulations. The developed model demonstrates its capacity to predict airfoil geometries with reasonable accuracy, effectively linking aerodynamic coefficients with corresponding shapes. A multihead self-attention block is incorporated as a feature-interaction module to process correlated aerodynamic input descriptors before mapping them to geometry parameters. The findings, derived from both the PARSEC and CST parameterization techniques, indicate that the framework reconstructs airfoil geometries with good agreement across diverse airfoil families within the considered Reynolds number and angle-of-attack envelope.
G. Eriş, Ahmet Can Özgören, O. Uzol· Journal of Aircraft· 0 citations
In the present work, an alternative methodology was developed for the rapid prediction of pressure distributions over wings of low-speed aircraft. A hybrid neural network architecture, named “MARTHA” (Model for Airloads Reconstruction using a Trained Hybrid Architecture), was presented, which is composed of a Multilayer Perceptron and the decoder of an Autoencoder. Three compact representation models—Principal Component Analysis (PCA), Autoencoder (AE), and Variational Autoencoder (VAE)—were systematically evaluated to determine the optimal dimensionality reduction architecture; the AE was selected based on its superior reconstruction accuracy and training stability. The main feature of MARTHA is that it provides predictions of the differential pressure coefficient field in the form of monochrome images, where the pixel intensity directly represents the normalized pressure value. One of the main objectives of developing MARTHA was to create a rapid surrogate model that can approximate vortex lattice method (VLM) simulations in preliminary design and optimization tasks, particularly when thousands of wing configurations need to be evaluated. The key feature of the proposed model is its ability to predict the pressure distribution for trapezoidal wings of various geometries 101–104 times faster than numerical models, while maintaining accuracy (R2 = 0.9998). The data obtained are presented in a convenient format for their further use in CAE systems of strength analysis. To assess the practical utility of the proposed model, implementation cases were carried out using the finite element software ANSYS 18.2 for three wing configurations not present in the training dataset. The pressure fields predicted by MARTHA were mapped onto the wing meshes, and linear static structural analyses were performed. The obtained Von Mises stress distributions showed good agreement with the corresponding distributions obtained using numerical models.
O. Lukyanov, Damian Josue Guerra Guerra, J. G. Quijada Pioquinto et al.· Technologies· 0 citations
Accurately modelling the nonlinear drag map within the design parameter space is a key challenge in preliminary nacelle design. However, due to sharp gradient variations in the drag distribution and the limited size of the available dataset, traditional surrogate models often suffer from insufficient predictive accuracy and poor generalization. To address this challenge, this study systematically evaluates the modelling performance of representative reduced-order models and deep learning–based approaches, and proposes a hybrid modelling framework (PMG) that integrates Proper Orthogonal Decomposition (POD), Multilayer Perceptron (MLP), and Gaussian Process Regression (GPR). The performance of various methods is evaluated and validated using high-resolution numerical simulations across the nacelle design parameter space. The results show that the PMG model reduces the required number of samples by 80% while accurately capturing the characteristics of complex drag distributions. Under small-sample conditions, the PMG model demonstrates superior predictive accuracy compared to traditional reduced-order models and deep learning-based approaches. This framework provides a promising approach for the rapid evaluation of preliminary nacelle designs.
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.