CCL

Research note

Journal of Chemical Theory and Computation

A Cross-Domain Graph Learning Protocol for Single-Step Molecular Geometry Refinement

Chengchun Liu, Wendi Cai, Boxuan Zhao, and Fanyang Mo

GeoOpt-Net uses SE(3)-equivariant, multifidelity learning to map inexpensive force-field conformers directly toward DFT-quality molecular geometries in one deterministic forward pass.

10.1021/acs.jctc.6c01080 View source code

01 / Problem

Problem

Reliable quantum-chemical calculations require accurate molecular geometries, but iterative DFT optimization becomes costly when many molecules must be processed.

02 / Approach

Approach

An SE(3)-equivariant graph architecture and multifidelity learning map an inexpensive initial conformer toward a target quantum-chemical geometry in one deterministic forward pass.

03 / Why it matters

Why it matters

A direct refinement step can reduce the geometry-preparation burden in large-scale quantum-chemical workflows while respecting the geometric symmetries of the task.

04 / Publication details

Publication details

Authors
Chengchun Liu, Wendi Cai, Boxuan Zhao, and Fanyang Mo
Journal
Journal of Chemical Theory and Computation
Year

BibTeX

@article{liu2026geooptnet,
  title   = {A Cross-Domain Graph Learning Protocol for Single-Step Molecular Geometry Refinement},
  author  = {Chengchun Liu and Wendi Cai and Boxuan Zhao and Fanyang Mo},
  journal = {Journal of Chemical Theory and Computation},
  year    = {2026},
  date    = {2026-07-12},
  volume  = {22},
  number  = {14},
  pages   = {7469--7479},
  doi     = {10.1021/acs.jctc.6c01080},
  url     = {https://doi.org/10.1021/acs.jctc.6c01080}
}