Exemplar-based objective classification of gust-induced loads across multiple flight conditions

This paper investigates whether an objective classification criterion can be found to organize the complexity of gust-induced loads across multiple flight conditions, while remaining as interpretable as labeling based on coarse parameters such as flight attitude. The authors state: "Is it possible to find an objective classification criterion that organizes the complexity of gust-induced loads across many flight conditions? And one that remains as interpretable as a labelling based on coarse parameters, such as the flight attitude?"

The motivation stems from the fact that small uncrewed aerial vehicles (UAVs) operate in low-altitude environments where the length and velocity scales of unsteady disturbances overlap directly with the vehicle dimensions, which makes gust-induced load excursions a critical threat to flight stability and structural integrity. Traditional gust models and linear frameworks fail in this regime, particularly for non-slender delta wings where the aerodynamics is dominated by leading-edge vortex (LEV) development and strongly nonlinear separation phenomena... under which linear superposition fails.

The authors propose an approach that encodes a large number of experimental observations through a machine-learned representation and applies a summarization procedure to select a minimal subset of highly significant exemplars. These exemplars, called textbook examples, provide a similarity-based objective classification criterion for all observations.

The method involves:

1. Training a Multi-Layer Perceptron (MLP) that takes four instantaneous pressure readings (Cp1 through Cp4) as inputs and predicts lift coefficient (CL) and side-force coefficient (CY) simultaneously as outputs

2. Using the output of the trained MLP truncated before its final linear layer as embedding coordinates (32-dimensional)

3. Applying k-medoids clustering in embedding space to select textbook subsets

4. Benchmarking against random subsets of equivalent size

The experimental facility consists of a 9×9 array of computer-controlled dual tube-axial DC fans capable of generating unsteady axial inflow. The aerodynamic model is a non-slender delta wing with a NACA0012 cross-section and a mid-span of chord c = 30cm. Four pressure taps are positioned near the leading edge, and a six-component force balance records time-resolved aerodynamic loads.

The database spans six vehicle attitudes, combining two angles of attack α ∈ 20°, 30° with three sideslip angles β ∈ 0°, 15°, 30°. The final database comprises 3479 gust events, partitioned into a training set of 2783 events and a test set of 696 events following an 80–20 split.

At the chosen textbook size of m* = 9, the textbook achieves a test mean-squared error of 0.161, compared to 0.473 for a random subset of the same size—a reduction of approximately 66%, while the full-database baseline stands at 0.099. This corresponds to a compression ratio of m*/m∞ ≈ 0.3%, representing a reduction of the full training set size by over two orders of magnitude.

The authors found that the learning curves for attitude-labeled and unlabeled textbooks do not show diverging trends to a significant level. This indicates that the summarization procedure already incorporates the essential data diversity, including in the attitude.

Nine fundamental response types were identified. The authors report: "Types 1, 9, 3 and 7 are exclusive to the two zero-sideslip attitudes (β = 0°) and occur at both angles of attack. Types 2, 5, 6, 8, 9 are associated with non-zero sideslip conditions, with 2 and 5 being the only types that only occur at a single attitude. Types 8 and 9 occur at four angles of attack and represent the most commonly occurring gust-load types. The authors conclude: The presence of cross-attitude clusters confirms that a subset of fundamental response types recurs independently of the flight configuration."

For response type 7, which produces the simultaneously strongest CL − CY responses, the authors found that around 90% of type 7 events occurs at α = 30. They note that type 7 load histories from α = 30, β = 0 are generally single-peaked and with a median duration of 108t* and median peak time at 53t*, while at α = 20, β = 0 they show generally longer median durations of 106t* and predominantly multi-peaked or sawtooth-shaped histories.

Drawing on existing PIV measurements from a previous study, the authors suggest: The PIVs exhibited leading-edge vortex (LEV) formation, followed by pinch-off which causes the flow to reattach and the lift force to revert to the baseline. They hypothesize that formation and pinch-off of isolated LEVs is more likely at higher α, while at α = 20, flow detachment appears to persist for longer and produce a series of weaker LEVs which form and pinch off intermittently.

The paper contributes three points of methodological novelty:

1. Isolating nine essential gust-load response types through a purely data-driven, objective procedure

2. Constructing an objective classification criterion by labeling observations according to their closest textbook exemplar, finding that most fundamental gust-load types are not confined to one attitude

3. Demonstrating that the manageable number of response types is advantageous for mapping unsteady flow physics, as a campaign of high-quality flow-field measurements could now be focused on the representative cases rather than distributed across the experimental parameter range

The authors conclude: This study provides a demonstration of a data-based framework for unfolding complex aerodynamic phenomena which goes beyond sheer predictive power and emphasizes simplicity and human-centric understanding.

Improvement 1: Interpretable Exemplar-Based Classification for Sparse Sensor Fusion

The AI system can be enhanced to select a minimal, human-interpretable subset of representative data points (e.g., textbook examples) from a high-dimensional sensor stream. Using a neural network’s penultimate layer as an embedding space, the system applies k-medoids clustering to identify exemplars that maximize classification accuracy while minimizing data storage. This enables real-time gust-load prediction on resource-constrained UAVs using only four pressure taps, with a 66% error reduction over random sampling and a 300× compression ratio. The system can autonomously flag novel aerodynamic states by measuring distance to the nearest exemplar, improving anomaly detection in unsteady flow regimes.

Improvement 2: Attitude-Agnostic Response Typing for Cross-Condition Generalization

The AI system can learn to cluster gust-load response types that are invariant to coarse flight parameters (e.g., angle of attack, sideslip). By benchmarking against attitude-labeled clusters and showing no significant divergence, the system can transfer learned response types across unseen flight conditions without retraining. This allows the AI to predict structural loads for new attitude combinations (e.g., α=25°, β=10°) by mapping to the nearest existing exemplar, reducing the need for exhaustive experimental campaigns. The system can also identify which response types are universal (e.g., types 8 and 9 appearing at four attitudes) versus attitude-specific, enabling adaptive control strategies that prioritize robust responses.

Improvement 3: Physics-Guided Temporal Pattern Mining for Flow Mechanism Inference

The AI system can extract physically meaningful temporal signatures from clustered load histories (e.g., single-peaked vs. sawtooth patterns) and correlate them with underlying flow phenomena like leading-edge vortex (LEV) formation and pinch-off. By linking cluster statistics (e.g., median duration, peak time) to prior PIV measurements, the system can generate hypotheses about flow physics—e.g., predicting that higher angles of attack favor isolated LEV pinch-off, while lower angles produce intermittent weaker vortices. This enables the AI to suggest targeted experimental measurements (e.g., PIV at specific exemplar conditions) to validate mechanisms, accelerating physics discovery and improving simulation fidelity for gust-load modeling.

Improvement 4: Minimal-Set Active Learning for High-Cost Measurement Prioritization

The AI system can use the textbook exemplars to guide an active learning loop, where expensive flow-field measurements (e.g., PIV or force balance) are only performed on the 9 representative gust events rather than the full 3479-event database. This reduces experimental cost by >99% while retaining predictive accuracy. The system can iteratively update exemplars as new data arrives, ensuring the minimal set remains representative of evolving flight envelopes. This is directly applicable to wind-tunnel testing, flight certification, and digital twin calibration for small UAVs operating in turbulent low-altitude environments.

Improved AI System Capabilities:

• Real-time gust-load prediction with 0.3% of training data, deployable on onboard processors.

• Cross-attitude load forecasting without retraining, enabling robust control across maneuvers.

• Autonomous discovery of flow regimes and generation of testable physics hypotheses.

• Cost-efficient experimental design by focusing high-fidelity measurements on canonical exemplars.

• Human-readable classification of aerodynamic states, facilitating pilot/engineer trust and regulatory compliance.