DDCCNet: Physics-Enhanced Multitask Neural Networks for Data-Driven Coupled-Cluster

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Machine Learning

Background:

• Coupled-cluster (CC) methods are highly accurate but computationally expensive for electronic structure calculations.
• Predicting CC amplitudes and energies from lower-level methods can significantly reduce computational cost.
• Deep learning offers a promising avenue for accelerating these predictions.

Purpose of the Study:

• To introduce the Data-Driven Coupled-Cluster Deep Network (DDCCNet), a novel deep learning architecture.
• To develop physics-enhanced deep learning models for predicting coupled-cluster singles and doubles (CCSD) amplitudes and correlation energies.
• To establish a scalable, physically grounded framework for efficient electronic structure prediction.

Main Methods:

• Developed three variants of DDCCNet (v1, v2, v3) with progressive architectural refinements.
• Incorporated parallel subnetworks for t1 and t2 amplitudes, feature-partitioned blocks, and physics-enhanced layers.
• Embedded symmetry and orbital-level interactions directly into the network structure for physical consistency.

Main Results:

• DDCCNet models jointly learn correlated amplitude patterns from lower-level electronic structure data.
• DDCCNet_v2 demonstrated the most accurate and transferable performance across diverse molecular systems.
• Achieved chemically precise correlation energies for methanol conformers, CO2 clusters, and small organic molecules.

Conclusions:

• DDCCNet provides a scalable and physically grounded framework for electronic structure prediction.
• The study successfully unifies machine learning and ab initio theory for efficient computations.
• DDCCNet represents a significant advancement in data-driven approaches to quantum chemistry.