Synergistic Multi-Task Learning for Electronic Density of States Prediction

First-principles calculations provide detailed electronic structure information but are computationally expensive, limiting their application to large-scale materials screening. Machine learning offers a promising alternative, yet existing approaches typically predict density of states (DOS), element-projected DOS (EPDOS), Fermi level, and band gap using separate models, missing potential synergies between these interrelated quantities. We propose DOSForge, a multi-task architecture that exploits such synergies via cross-conditioning between decoders and parameter-free band gap supervision. Experiments show consistent gains across all four targets (DOS, EPDOS, Fermi level, and band gap), whereas naive multi-task baselines exhibit clear trade-offs across targets. DOSForge achieves state-of-the-art DOS prediction and demonstrates that principled multi-task design can turn competing objectives into mutual gains.