Experimental Design for Aerodynamic Dataset Modelling using Bayesian Neural Networks and BALD Scoring

Bayesian Experimental Design (BED) provides a practical framework for selecting an experimental design that maximizes the expected usage. Within the scope of the present work, we apply the BED paradigm to high-dimensional aerodynamic stability and control datasets. A Bayesian Neural Network (BNN) serves as a surrogate model of an initial dataset, approximating the posterior distribution via mean-field Variational Inference (VI). The inherent uncertainty quantification capabilities of the BNN enable the decomposition of uncertainty into aleatoric and epistemic components utilizing a dual-head architecture. Leveraging this, the Bayesian Active Learning by Disagreement (BALD) method provides a scoring criterion to estimate the Expected Information Gain (EIG) of hypothetical, unlabeled data points. The candidate pool is generated using a Sobol sequence. In the context of Aerodynamic Dataset Modelling (ADM), a prudent design of new experiments is driven by economic constraints. Data acquisition techniques predominantly focus on Wind-Tunnel Testing (WTT), supplemented by numerical simulations and flight testing. The proposed methodology aims to iteratively identify the subsequent optimal measurements, thereby optimizing the test matrix setup. A greedy plus diversity strategy ensures comprehensive coverage of the input space while maintaining a high EIG. To evaluate the performance and applicability of the implemented methods, data from WTT campaigns of the DLR-F17 and DLR-F19 Unmanned Aerial Vehicles (UAVs), commonly known as SACCON, is utilized. This configuration features a highly swept, low-observable lambda wing planform without a separate empennage, representing a high-agility aircraft configuration. Results demonstrate that the introduced approach reliably identifies potential reductions in data acquisition needs, thus decreasing costs and resource demands of WTT campaigns, numerical computations, or flight testing, while maintaining the predictive accuracy of the BNN surrogate model.     «

Bayesian Experimental Design (BED) provides a practical framework for selecting an experimental design that maximizes the expected usage. Within the scope of the present work, we apply the BED paradigm to high-dimensional aerodynamic stability and control datasets. A Bayesian Neural Network (BNN) serves as a surrogate model of an initial dataset, approximating the posterior distribution via mean-field Variational Inference (VI). The inherent uncertainty quantification capabilities of the BNN ena...     »