PaperlayerView exact revision on arXiv

Source-linked AI summary

Improving aircraft performance using machine learning: a review

Soledad Le Clainche, Esteban Ferrer, Sam Gibson, Elisabeth Cross, Alessandro Parente, Ricardo Vinuesa

arXiv:2210.11481v1cs.LGphysics.data-anphysics.flu-dyn

TL;DR

Aerospace engineering needs methods that improve performance and sustainability while handling costly, data-intensive simulations and experiments. This review synthesizes ML concepts, strategies, and applications across aerospace disciplines, finding broad potential to improve aircraft performance and influence the field. It also identifies data-fidelity and modelling limitations that constrain deployment.

  • Problem

    Aerospace research must address environmental impact, aircraft performance, and costly high-fidelity simulations and experiments.

  • Method

    The review synthesizes ML concepts, strategies, advantages, challenges, and applications across aerospace engineering disciplines.

  • Results

    The review reports that ML is improving aircraft performance and has broad applications in fluid dynamics, aerodynamics, acoustics, combustion, flight testing, and structural health monitoring.

  • Takeaways & Limitations

    ML should improve efficiency and robustness within CFD subproblems rather than replace CFD, while supporting models and databases for cleaner, safer aircraft designs.

  • Takeaways & Limitations

    Some ML approaches require high-fidelity training data, limiting applicability beyond the cases represented during training.

Abstract

from arXiv · show

This review covers the new developments in machine learning (ML) that are impacting the multi-disciplinary area of aerospace engineering, including fundamental fluid dynamics (experimental and numerical), aerodynamics, acoustics, combustion and structural health monitoring. We review the state of the art, gathering the advantages and challenges of ML methods across different aerospace disciplines and provide our view on future opportunities. The basic concepts and the most relevant strategies for ML are presented together with the most relevant applications in aerospace engineering, revealing that ML is improving aircraft performance and that these techniques will have a large impact in the near future.

1 Introduction

Aerospace engineering faces sustainability, safety, cost, and competitiveness challenges that demand new technologies. This review presents ML as an approach for exploiting aerospace data and supporting applications across aircraft performance disciplines.

  • Sustainability goals: 2050 targets include 90% lower NOx, 75% lower CO2 per passenger/kilometre, and 65% lower aircraft noise than typical 2000 capabilities.
  • Safety goals: Aviation must also reduce accidents, with a stated goal of 80% fewer accidents by 2050 than in 2000 despite rising traffic.
  • Research challenge: High-fidelity experiments and simulations provide valuable aerodynamic, combustion, and flow-control data, but their economic and computational costs motivate alternatives.
  • Review scope: ML applications in aerospace include data interpretation, system management, aircraft modelling, reduced-order modelling, forecasting, flow control, acoustics, combustion, flight testing, and structural health monitoring.
  • Review scope: The review surveys ML strategies and applications across fluid dynamics, aerodynamics, acoustics, combustion, and solid mechanics to identify research opportunities for cleaner and safer aircraft.

2 Machine learning methodology: a general overview

ML extracts information from increasingly large aerospace databases and uses it to build models that support sustainable aircraft development. The section introduces ML methods and their aerospace relevance.

  • Aerospace databases contain information that ML algorithms can exploit to gain physical insight and develop cleaner, safer aircraft designs.
  • ML is used to develop reduced-order and surrogate models from complex aerospace data.
  • Figure 2 provides a general classification of ML methods and identifies techniques most popular in aerospace engineering.

2.1 Neural networks

Neural networks map inputs through parameterized layers and optimize those mappings during training. Their architecture, activation functions, and regularization determine how well they represent and generalize aerospace data.

  • ML uses artificial neural networks to process databases and solve optimization problems through back propagation and stochastic gradient descent.
  • Regularization constrains the highly under-determined network optimization problem and helps prevent overfitting so solutions generalize to different data sequences.
  • A multilayer neural network maps an input X through successive functions and weight matrices to produce an output Y.
  • Network performance depends on layer dimensions, connections, mappings, and activation functions, including linear, sigmoid, tanh, and ReLU functions.
  • RNNs process sequences through recurrent connections, while CNNs are widely used for ML applications in fluid dynamics.

2.2 Regression and Classification

Regression and classification are supervised-learning tasks that learn from labeled data, while related methods extend prediction and decision-making to different data structures. Their usefulness depends on data quality, model choice, and computational cost.

  • Supervised learning: Regression approximates continuous outputs, whereas classification assigns inputs to discrete categories using labeled examples.
  • Regression: Linear regression fits relationships such as y = mx + c and can interpolate missing data or extrapolate beyond the training-data range.
  • Classification: Logistic regression uses a sigmoid curve and a 50% threshold to perform binary classification.
  • Classification: Decision trees classify through sequential branch decisions, while random forests combine multiple decision trees for classification or regression.
  • Classification: K-nearest neighbours relies on distances to classified neighbours and works best when datasets are not excessively large and are free of noise.
  • Classification and regression: Support-vector machines handle classification and regression in large transformed feature spaces but incur high computational cost on large datasets.
  • Probabilistic regression: Gaussian-process regression represents predictions as a distribution of functions conditioned on data and is widely used in structural health monitoring.

2.3 Semi-supervised learning

Semi-supervised aerospace ML includes GANs for generating high-resolution outputs from low-resolution inputs and deep reinforcement learning for learning actions that maximize cumulative reward.

  • Generative adversarial networks: GANs combine a generator and discriminator trained against each other using competing tasks.In super-resolution, the generator produces a high-resolution output from a low-resolution input consistent with the reference image’s statistical features.
  • Generative adversarial networks: GAN-based super-resolution can increase the resolution of quantities measured at the wall in turbulent channel flow.The architecture includes convolution, activation, normalization, and sub-pixel convolution layers.
  • Deep reinforcement learning: Deep reinforcement learning uses an agent that observes an environment, applies actions, receives rewards, and iteratively changes the environment state.For flow control, the environment is the flow and the agent can be a neural network.
  • Deep reinforcement learning: The objective of deep reinforcement learning is to find a decision policy that maximizes cumulative reward.Different implementations should be selected according to the dynamics of the problem.

2.4 Unsupervised learning: clustering and dimensionality reduction

Unsupervised learning discovers structure in unlabeled aerospace data through clustering and dimensionality reduction, while methods such as local PCA and kernel PCA address nonlinear relationships with different computational trade-offs.

  • Unsupervised learning: Unsupervised learning processes unlabeled data to identify structure and relevant features for lower-dimensional representations.This makes it relevant to aerospace applications involving large datasets.
  • Clustering: Clustering methods discover groups without pre-sampled data, including K-means, mean shift, DBSCAN, and Gaussian-mixture models.K-means iteratively assigns points to the nearest centroid and updates centroids until they stop moving; mean shift instead seeks dense regions.
  • Clustering: ML clustering outperformed classical viscous sensors for identifying boundary-layer and wake regions in RANS simulations.Gaussian-mixture methods also identify laminar and turbulent regions in LES and can support local mesh adaptation.
  • Dimensionality reduction: PCA represents data with an orthogonal basis and retains principal components that account for most variance, producing a low-dimensional linear predictor.The components maximize variance rather than interpretability, which can complicate physical interpretation.
  • Local dimensionality reduction: Global PCA is limited for nonlinear aerospace flow data, motivating local approaches that approximate the manifold with locally linear regions.VQPCA can use supervised or unsupervised clustering before applying PCA within clusters.
  • Kernel PCA: Kernel PCA maps data into a high-dimensional feature space where nonlinear relationships can be treated linearly through a kernel function.Its covariance matrix scales with observations, and reconstructing data from scores requires solving a minimization problem.
  • Dimensionality reduction: Autoencoders reduce dimensionality with an encoder that learns codes and a decoder that reconstructs the data while minimizing reconstruction error.They use neural networks with linear or nonlinear activation functions to extract principal data features.

3 Fluid mechanics

Machine learning is being applied across fluid mechanics to accelerate simulations, reduce model dimensionality, improve turbulence prediction, and augment experimental measurements. These methods show promise, but some approaches remain constrained by the need for high-fidelity training data.

  • Computational fluid dynamics: ML can accelerate CFD, improve turbulence modeling, and support efficient numerical treatments such as Poisson resolution and reduced computational domains.Neural networks have also been used to add corrective forcing that preserves high accuracy in lower-order simulations.
  • Computational fluid dynamics: Deep-learning subgrid-scale models for LES can reproduce high-fidelity behavior, but their applicability is limited by the need for high-fidelity training data.This dependence restricts generalization beyond cases represented during training.
  • Computational fluid dynamics: Reinforcement learning offers an unsupervised alternative for LES subgrid-scale modeling that does not require high-fidelity reference data and can target consistency with the coarse simulation.This approach may favor generalizability because it need not emulate a high-fidelity simulation exactly.
  • Reduced-order models: Reduced-order models simplify high-dimensional, chaotic turbulent systems for analysis, flow control, and optimization by separating or compressing their spatial and temporal structure.POD ranks spatial modes by energy contribution, while DMD ranks modes by frequency; HODMD incorporates nonlinear temporal dependencies.
  • Reduced-order models: ML models can predict turbulent-flow temporal dynamics, including turbulence statistics, Poincaré maps, Lyapunov exponents, and extreme events.Reported approaches include LSTM, Koopman-based models, physics-constrained reservoir computing, deep learning, and autoencoders.
  • Experiments: ML can augment experiments by predicting unmeasured quantities from wall measurements, including heat flux and velocity fluctuations away from the wall.Wall shear stress, wall pressure, and convolutional neural networks provide non-intrusive sensing routes for reconstructing flow behavior.

4 Aerodynamics

Machine learning supports aerodynamic modeling and design by providing surrogate and reduced-order models for coefficient estimation, unsteady responses, aeroelasticity, and optimization. The review also identifies continuing needs for data fusion, physics-informed modeling, interpretability, and broader validation in complex flows.

  • Aerodynamic modeling: Aerodynamic data from CFD and experiments includes distributed quantities such as pressure and skin friction and integral quantities such as drag, lift, and moment coefficients.The high cost of obtaining these data motivates data-driven aerodynamic models.
  • Aerodynamic modeling: Data-driven aerodynamic models include semi-empirical models, surrogate models, and reduced-order models for representing flow dynamics or predicting outputs from geometry and flow conditions.These model classes support aerodynamic analysis and design tasks.
  • Aerodynamic modeling: Neural networks provide flexible tools for reduced-order and surrogate modeling, enabling faster CFD-based evaluation and aerodynamic optimization across candidate geometries.Their applications include estimating integral and distributed aerodynamic quantities for multiple designs.
  • Future directions: The review highlights data fusion across simulations with different fidelities and experiments, alongside models that incorporate flow physics and improve interpretability.These needs become more important for turbulent, three-dimensional, multiscale, and industrial flows.
  • Aerodynamic coefficients estimation: A neural-network surrogate trained on about 100,000 CFD cases estimated transport-aircraft aerodynamic coefficients about 4000 times faster than the CFD solver and supported optimization design.Inputs included flight condition, airfoil geometry, and wing planform.
  • Aeroelasticity: Neural networks were used to estimate nonlinear aeroelastic behavior and aerodynamic forces across multiple flow conditions, including transonic pitching and plunging cases.LSTM models captured time-delayed effects in unsteady aerodynamics and were tested on NACA 64A010 configurations.
  • Design optimization: Surrogate models combined with optimization algorithms can select airfoil designs for transonic conditions while minimizing time-averaged drag and reducing buffet effects.The stated objective includes preventing structural damage associated with buffet.
  • Scope and challenges: ML-based aerodynamic models span incompressible to transonic compressible-flow problems, while their broader use remains tied to model accuracy, data coverage, physics integration, and complex-flow challenges.The review identifies aerodynamic-force estimation, aeroelastic response prediction, and design optimization as major application areas.

5 Aeroacoustics

Aeroacoustics couples turbulent-flow noise generation with acoustic propagation, making large simulations costly and motivating hybrid models and ML surrogates. The review finds substantial potential for ML, but applications remain limited and require caution for non-smooth transfer functions.

  • Aeroacoustics must account for turbulent flow interacting with geometry, acoustic generation, and subsequent propagation, unlike simpler acoustic problems.
  • Hybrid aeroacoustic approaches reduce computational cost by separating highly accurate near-field flow calculations from coarser far-field acoustic propagation.
  • ML models, including neural networks, can map flow or geometric inputs to sound pressure levels and predict acoustic-wave propagation.
  • Existing ML applications in acoustics remain scarce and are largely restricted to automotive problems or simple flows.
  • Aeroacoustic ML has substantial potential, but non-smooth transfer functions can make neural networks unreliable for derivative-based tasks such as gradient or adjoint optimization.

6 Combustion

Combustion research uses ML and reduced-order methods to analyze high-dimensional turbulent reacting flows, reduce simulation costs, and improve predictive modeling. The review highlights gains in feature extraction, chemical-mechanism optimization, sub-grid modeling, and fast system-response prediction, alongside interpretability and computational challenges.

  • Combustion research handles massive experimental and simulation datasets, motivating dimensionality reduction and data-driven analysis of turbulent reacting flows.
  • Data analysis and feature extraction: Neural-network regression identified features of MILD combustion and predicted coexistence of combustion modes, supporting adaptive closure selection in simulations.
  • Data analysis and feature extraction: Nonlinear dimensionality-reduction methods can improve reduction potential but may become computationally intractable, require many hyperparameters, and produce difficult-to-interpret features.
  • Dimensionality reduction, classification and adaptive chemistry: Optimization methods address inverse chemical-mechanism problems and improve predictions using increasingly available experimental data, including mechanisms for complex fuels.
  • Combustion modeling: Data-driven methods support improved sub-grid and turbulence closures, while deep CNNs are used to predict unresolved combustion quantities.
  • Reduced-order models for realistic combustion systems: Reduced-order models approximate hidden input-output relationships for fast control and optimization, but black-box approaches lack the guarantees of physics-based methods.

7 Structural assessment

Aerospace structural assessment must address harsh loading, difficult-to-predict composite failure, costly inspection, and sensor-network constraints. ML methods—including probabilistic virtual sensing, physics-informed Gaussian processes, and grey-box models—support damage detection, load prediction, fatigue analysis, and risk-informed maintenance.

  • Aerospace structures face broad loading spectra and harsh environments, while composite materials introduce more difficult-to-predict failure modes.
  • Structural health monitoring can reduce aircraft downtime and maintenance costs while improving safety and reliability, especially where inspection is difficult and failure is costly.
  • Complex and heavy sensor wiring can be prohibitive, motivating low-power or wireless networks and ML methods for interpreting sensor data and detecting faults.
  • Embedding physical insight in Gaussian-process kernels gives posterior samples a physical interpretation, while grey-box input augmentation incorporates known structural processes.
  • Gaussian-process regression predicts aircraft loads and stresses while its posterior uncertainty can be propagated through stress-life analysis to estimate probabilistic fatigue damage.
  • Probabilistic fatigue approaches can support risk-based decisions, improve risk control, and enable more effective maintenance planning for aerospace components.

8 Conclusions

ML is broadening aerospace applications across fluid dynamics, aerodynamics, acoustics, combustion, and structural health monitoring, while improving simulation efficiency and physical insight. However, data-source limitations and the lack of robust, generalizable schemes remain open challenges.

  • ML tools now address problems across fundamental fluid dynamics, aerodynamics, acoustics, combustion, and structural health monitoring.
  • Data fusion can balance model accuracy against data-generation cost by combining experimental and numerical data at different fidelities.
  • Most industrial models still rely on a single data source, leaving opportunities for broader multi-source approaches.
  • Robust and generalizable schemes remain an open research topic despite promising performance from physics-grounded models.
  • Future ML development should improve CFD solver subproblems rather than replace CFD, including simulation acceleration, enhanced LES and RANS modeling, and reduced-order modeling.
Loading 2210.11481v1…