Adaptive hierarchical feature extraction with transformer-augmented CNNs for high-dimensional fluid mechanics data analysis

We propose an adaptive hierarchical feature extraction framework for high-dimensional fluid mechanics data analysis, combining multi-scale convolutional neural networks with transformer-based temporal modeling to address the challenges of processing complex spatiotemporal flow fields. The proposed method integrates parallel CNN branches with dilated convolutions to capture multi-resolution spatial features, then dynamically fuses them through an attention mechanism that adaptively weights scales based on local flow characteristics. The architecture incorporates physics-informed normalization to maintain dimensional consistency with conventional computational fluid dynamics data structures, enabling seamless integration with existing numerical solvers and experimental measurements. Key innovations include the learnable multi-scale feature fusion strategy, the hybrid CNN-transformer design for joint spatial-temporal analysis, and the physics-aware interface that preserves interpretability. Experimental validation demonstrates superior performance in flow reconstruction and regime classification tasks compared to conventional deep learning approaches, particularly for flows exhibiting multi-scale phenomena such as boundary layer separation and vortex shedding. The framework provides a versatile tool for analyzing high-dimensional fluid mechanics data while maintaining computational efficiency and interpretability.