Redefining the data foundations of turbulence research: a physics-informed active sampling approach

Poorly organized flight-state data remain a major obstacle in the reliable prediction of aviation turbulence, as instability-prone flight conditions are often buried among abundant routine operations. When these rare but critical states are obscured, predictive models fail to recognize the aerodynamic patterns that precede turbulence, reducing sensitivity to real-world hazardous events. This study explores whether combining uncertainty-based active sampling with physics-informed selection criteria can yield more coherent and structurally organized datasets, in which physically distinct rare states emerge as well-represented groups separate from routine operational data, thereby enhancing both dataset diversity and the physical interpretability of data groupings. Flight data were processed through a dual-layer sampling framework that assigns adaptive weighting to statistical uncertainty and aerodynamic relevance, and selected datapoints were represented in a multidimensional feature space using k-means clustering to evaluate improvements in data organization. Results showed that this combined approach produced clearer separation between physically distinct flight regimes and yielded higher clustering quality metrics than conventional sampling. Beyond these results, the core contribution of this work is its pioneering nature—it introduces machine-driven active sampling with physics-informed reasoning at the data-preparation stage, a direction rarely explored in aerospace research. The scope of this study is specifically focused on evaluating whether combining uncertainty-based and physics-informed sampling can improve dataset structural quality, as measured through clustering coherence metrics (e.g., silhouette score, Calinski-Harabasz index, Davies-Bouldin index), rather than directly assessing downstream predictive model performance. This study should therefore be regarded as a proof-of-concept that examines the potential of adaptive, physics-aware data curation strategies to enhance the intrinsic organization and coherence of turbulence datasets, establishing a foundation for future work on turbulence modeling.