Physics-Aware Representation Learning on Electronic Charge Density for Materials Property Prediction

The fundamental quantity governing the mechanical and thermodynamic properties of a crystalline solid is its electronic charge density. Yet, its direct use for the rapid prediction of materials properties remains challenging due to its high dimensionality. Here, we present a physics-informed deep learning framework that directly predicts mechanical and thermodynamic properties from the three-dimensional electronic charge density derived from density functional theory (DFT). The proposed approach first utilizes a three-dimensional convolutional autoencoder for unsupervised dimensionality reduction, compressing a high-resolution charge-density grid (128 × 128 × 128) into a compact latent representation (16 × 16 × 16 × 16) while preserving physically meaningful features, as confirmed by negligible reconstruction errors across diverse crystal systems. The compressed latent-space representation of charge density is then used by two different regression models for property prediction: Light Gradient Boosting Machine (LightGBM) and Attention-based 3D Convolutional Neural Networks (Att CNN), and their performance is compared. Combining composition-based descriptors (Material Agnostic Platform for Informatics and Exploration or MAGPIE) with electronic charge density data further improves the model accuracy. Using a data set of about 6059 inorganic compounds spanning multiple crystal symmetries, the models achieve strong predictive performance for bulk modulus K (R² = 0.94), Young's modulus E (R² = 0.88), shear modulus G (R² = 0.87), formation energy E_form (R² = 0.96), and Debye temperature Θ (R² = 0.89). This work establishes electronic charge density as a transferable, physics-grounded descriptor for materials property prediction, requiring ≈1/25 of the computational resources of full-fledged DFT calculations.