DDCCNet: Physics-enhanced Multitask Neural Networks for Data-driven Coupled-cluster

TL;DR

DDCCNet introduces physics-enhanced multitask neural networks that accurately predict coupled-cluster energies, integrating physical principles into deep learning for efficient electronic structure calculations across diverse molecules.

Contribution

The paper develops a novel multitask deep learning framework, DDCCNet, that incorporates physical insights to improve the prediction of coupled-cluster amplitudes and energies.

Findings

DDCCNet_v2 achieves high accuracy across molecular systems.

Models incorporate symmetry and orbital interactions.

Scalable framework unifying machine learning and ab initio methods.

Abstract

We present the data-driven coupled-cluster deep network (DDCCNet), a family of multitask, physics-enhanced deep learning architectures designed to predict coupled-cluster singles and doubles (CCSD) amplitudes and correlation energies from lower-level electronic structure methods. The three DDCCNet variants (termed as v1, v2, and v3) progressively incorporate architectural refinements ranging from parallel subnetworks for t_1 and t_2 amplitudes to feature-partitioned blocks and physics-enhanced intermediate prediction layers that are structured in accordance with coupled-cluster equations to enhance physical consistency and multitask learning efficiency. These models jointly learn correlated amplitude patterns while embedding symmetry and orbital-level interactions directly into the network structure. Applied to methanol conformers, CO2 clusters, and small organic molecules, DDCCNet_v2 delivered the most accurate and transferable performance, achieving chemically precise correlation energies across diverse molecular systems. Collectively, DDCCNet establishes a scalable, physically grounded framework that unifies machine learning and ab initio theory for efficient, data-driven electronic structure prediction.