An interpretable multi-scale framework for unsupervised feature extraction in hierarchical 1D data

In many scientific domains involving one-dimensional data processing, such as spectral analysis, classical models like Principal Component Analysis (PCA) are favoured for their interpretability but fail to capture nonlinear structures. Conversely, the ‘black box’ nature of deep learning impedes its adoption where trust is critical. To address this, we propose an explainable AI framework featuring the Multi-Scale Decomposition Mixing Autoencoder (MDM-AE). Its novel ‘top-down hierarchical mixing strategy’ is physics-informed, fusing the data’s physical context to learn a geometrically superior latent space from its intrinsic manifold. To ensure transparency, we introduce and improve upon ‘decoder-weighted Grad-CAM,’ an interpretability method that provides scientifically consistent visual explanations for the model’s feature extraction process. Evaluated on public spectral datasets, MDM-AE demonstrates significantly superior unsupervised clustering quality over baselines like PCA and standard autoencoders. Our interpretability analysis confirms that this performance stems from an adaptive, chemically sound feature selection strategy. This framework surpasses classical model performance while providing the transparency required for trustworthy scientific discovery from unlabelled sequential data.