Learning disentangled latent representations facilitates discovery and design of functional materials

The discovery of new materials is often constrained by the need for large labelled datasets or expensive simulations. In this study, we explore the use of Disentangling Autoencoders (DAEs) to learn compact and interpretable representations of spectral data in an entirely unsupervised manner. We demonstrate that the DAE captures physically meaningful features in optical absorption spectra, relevant to photovoltaic (PV) performance, including a latent dimension strongly correlated with the Spectroscopic Limited Maximum Efficiency (SLME)--despite being trained without access to SLME labels. This feature corresponds to a well-known spectral signature: the transition from direct to indirect optical band gaps. Compared to Principal Component Analysis (PCA) and a beta-Variational Autoencoder (beta-VAE), the DAE achieves superior reconstruction fidelity, improved correlation with efficiency metrics, and more compact encoding of relevant features. We further show that the DAE latent space enables more efficient discovery of high-performing PV materials, identifying top candidates using fewer evaluations than both VAE-guided and random search. These results highlight the potential of DAEs as a powerful tool for unsupervised structure-property learning and suggest broad applicability to other areas of materials discovery where labeled data is limited but rich structure is present in raw signals.