QGeoSEP: A Novel Multi‐Task Learning Framework Integrating Quantum, Geometric, and Semantic Features for Collaborative Prediction of Multiple Properties With Potential Application to Energetic Materials

ABSTRACT Accurate physicochemical property prediction is critical for the rational design of energetic materials (EMs), yet limited high‐quality experimental data and property‐wise data imbalance restrict the application of conventional single‐task and multi‐task machine learning models. Here, we propose QGeoSEP, a multi‐task learning framework for the multi‐property prediction of EMs under data imbalance and scarce annotation. Integrating three‐dimensional geometric structures, electronic structure, features and endpoint semantic information, QGeoSEP captures attribute interdependencies via the transformer's multi‐head attention mechanism. It outperforms single‐task baselines and the multi‐task model MTL‐GNN on the QM9 dataset. On a curated dataset of 211,890 EMs‐like C–H–O–N organic small molecules, it achieves R 2 of 0.949, 0.878, 0.891, and 0.647 for density, melting point, heat of combustion, and decomposition temperature, respectively, with a 14.37% average improvement over MTL‐GNN. On an independent 661‐EMs test set, it yields superior prediction accuracy for density and decomposition temperature, with MAE of 0.090 g·cm −3 and 21.641 K. QGeoSEP trained on the C–H–O–N dataset addresses EMs data scarcity, enables collaborative multi‐property prediction under data sparsity and imbalance, and exhibits excellent transfer generalization for real EMs. The model is deployed on a web platform for barrier‐free use without programming or local installation.