Density Matrix-Driven Multi-Task Learning for Coupled-Cluster Accuracy

Shao et al. (2023) and Tang et al. (2024) both use ML for electronic structure, but Shao focuses on density matrices while Tang targets CCSD(T) accuracy. This idea merges them: train a single multi-task model to predict both the 1-RDM and CCSD(T)-level observables (energies, forces, dipole moments). The 1-RDM acts as a physically interpretable intermediate, ensuring consistency across outputs. Unlike Tang’s property-only model, this enables downstream tasks (e.g., excited states via TD-DFT on the predicted 1-RDM). It leverages density functional theory’s bijective maps (Shao) while surpassing DFT accuracy, potentially democratizing CCSD(T) for materials (Wei et al. 2024) and biomolecules.

References:

1. Scalable Ab Initio Electronic Structure Methods with Near Chemical Accuracy for Main Group Chemistry.. Yujing Wei, Sibali Debnath, John L. Weber, Ankit Mahajan, D. Reichman, R. Friesner (2024). Journal of Physical Chemistry A.
2. Machine learning electronic structure methods based on the one-electron reduced density matrix. Xuecheng Shao, Lukas Paetow, M. Tuckerman, M. Pavanello (2023). Nature Communications.
3. Machine learning electronic structure methods based on the one-electron reduced density matrix. Xuecheng Shao, Lukas Paetow, M. Tuckerman, M. Pavanello (2023). Nature Communications.
4. Multi-task learning for molecular electronic structure approaching coupled-cluster accuracy. Hao Tang, Brian Xiao, Wenhao He, Pero Subasic, A. Harutyunyan, Yao Wang, Fang Liu, Haowei Xu, Ju Li (2024). arXiv.org.
5. Machine learning electronic structure methods based on the one-electron reduced density matrix. Xuecheng Shao, Lukas Paetow, M. Tuckerman, M. Pavanello (2023). Nature Communications.