The search for the gust-wing interaction "textbook"

We address whether complex physical relations can be investigated through the synergy of automated high-volume experiments and the reduction of large datasets to a concise, representative subset of canonical examples – a “textbook”. To this end, we consider the unsteady aerodynamics of wing-gust interactions, which is characterized by its rich, high-dimensional physics. We take advantage of a purpose-built gust generator to systematically produce over 1,000 distinct random gust events and to measure the unsteady loads induced on a delta wing. We then employ a data summarization procedure to identify representative subsets of increasing size from the large-scale database, which then serve as training data for a machine-learning model of the aerodynamic loads from sparse pressure measurements. An appropriately selected “textbook” of a few events can achieve predictive accuracy comparable to random training sets up to two orders of magnitude larger, capturing the intrinsic diversity of the full-scale data and enhancing modeling efficiency and interpretability. Our methodology evidences the potential of distilling the essential information contained in large amounts of experimental observations.

💡 Research Summary

**
The paper investigates whether complex physical relationships can be efficiently captured by combining automated high‑throughput experiments with a systematic reduction of large datasets to a concise, representative subset—a “textbook”. The authors focus on the unsteady aerodynamics of gust‑wing interactions, a problem characterized by high dimensionality and a wide variety of extreme and edge cases.

A purpose‑built gust generator consisting of an 81‑fan array produces random axial gusts by independently varying three forcing parameters: base fan velocity, velocity increment, and non‑dimensional forcing interval. Over a series of randomized runs, more than 1 000 distinct gust events are generated. Each event lasts one minute and provides time‑resolved data from four pressure taps embedded in a non‑slender delta wing (NACA 0012 cross‑section) and six‑component force/moment measurements from a balance. The pressure signals are nondimensionalized (2 U₀² ρ) and used as inputs, while the lift coefficient CL (2 Fy / c U₀² ρ) serves as the output. After segmentation, the final database contains 1 031 events with a median duration of roughly 110 convective time units, equivalent to about 13 000 convective times of flight.

For load prediction, a simple feed‑forward neural network (Multi‑Layer Perceptron) is employed. The network has four input neurons (the four pressure coefficients), four hidden layers of 16 neurons each, PReLU activations, and a single output neuron for CL, totaling 977 trainable parameters. An 80/20 split yields 824 training and 207 test events. The model is trained by minimizing mean‑squared error (MSE) and hyper‑parameters are selected via grid search.

Two numerical experiments assess the value of individual training samples. First, 20 events are each used alone to train separate MLP instances; each model is then evaluated on the remaining 19 events. The results reveal substantial variability: some events (e.g., #6) consistently produce low test errors across many unseen cases, while others (e.g., #20, #8) lead to high errors, indicating that not all data are equally informative. Second, learning curves are generated by training the MLP on random subsets of increasing size m. The test MSE drops sharply up to m≈400–600 and then plateaus, showing that a fraction of the full dataset already yields near‑optimal generalization.

The central contribution is a data‑summarization procedure that extracts a “textbook” subset D_txt ⊂ D. Using distance‑based clustering (k‑centers) combined with model‑based importance metrics, the authors select a small number of events that preserve the diversity of the full manifold, including edge and extreme cases. Textbooks containing as few as 10–30 events achieve predictive performance comparable to random training sets that are two orders of magnitude larger, while dramatically reducing training time and memory requirements.

The findings demonstrate that (1) large experimental databases can be compressed without losing essential physics, (2) a carefully chosen, physically diverse subset suffices for high‑accuracy machine‑learning models, and (3) such compression improves interpretability and makes real‑time deployment feasible for autonomous flight systems with limited sensing and computational resources. Moreover, because the textbook selection is not tied to a specific learning architecture, the approach can be extended to other fluid‑structure problems, offering a general pathway toward data‑efficient scientific discovery.