Bringing physical insight into machine learning through COGITO-based architectures

Machine learning has become a powerful tool in materials design, yet most approaches operate as black boxes—sacrificing physical interpretability for accuracy. Here, we introduce a framework that bridges this divide by coupling Crystal Orbital Guided Iteration To atomic-Orbitals (COGITO) with machine learning. COGITO transforms plane-wave DFT outputs into atom- and bond-resolved representations of electronic structure, providing physically meaningful features for model training. We demonstrate that this physics-guided pipeline enables predictive and interpretable learning of the density matrix, charge transfer, and electronic structure. Moreover, by linking ML predictions directly to the underlying atomic orbitals, COGITO enables rational control over chemistry and structure to achieve targeted electronic properties. This approach opens a path toward explainable AI for materials—where learning and design are guided by the quantum-mechanical origins of bonding.