Tri-Module Deep DFT Architecture with Physical Regularization, Task Coupling, and Compression-Based Transferability

We introduce a deep learning architecture designed to enhance Density Functional Theory (DFT) property prediction through a tri-module system combining task decoupling, physics-based regularization, and compression-based transferability. The proposed framework consists of a shared backbone followed by three dedicated heads for energy, force, and polarizability predictions. A physically inspired Jacobian regularizer enforces structural consistency across outputs, while a bottleneck-based latent space is used to compress task-relevant features and enable inter-task knowledge transfer. Our design allows for minimal interference between learning objectives while preserving physical alignment, especially under low-data or high-noise conditions. This architecture aims to improve generalization, interpretability, and adaptability in multi-task DFT-based modeling without compromising computational efficiency. Early results suggest robust performance across benchmarks and a scalable foundation for future cross-property extensions.