OrbFlow predicts GTO coefficients with SE(3)-equivariant flow matching, cutting density error on QM9 by 13.6% and SCF iterations by up to 68%

OrbFlow predicts coefficients directly on a compact atom-centered GTO basis while learning a probability path over the entire coefficient space via flow matching, replacing pointwise regression. Grid-based architectures were computationally costly, while basis-set methods struggled to capture structural correlations in coefficient space; this work brings generative modeling into coefficient space, combining the efficiency of a compact basis with correlation modeling. The abstract reports state-of-the-art accuracy on QM9, with density error reduced by 13.6% relative to the previous best model.

A two-phase trajectory curriculum is used to mitigate discretization drift during numerical integration. This is a training-strategy design aimed at error accumulation when integrating in coefficient space, rather than only changing the network architecture. The abstract describes the curriculum as part of the training method and states it alongside the final accuracy gains.

The predicted density serves as an SCF initialization that substantially reduces iteration counts and supports zero-shot transfer to unseen exchange-correlation functionals. Unlike prior surrogates focused mainly on density accuracy, this work plugs density prediction directly into the SCF procedure and quantifies the speedup. The abstract reports SCF iterations cut by up to 68% and zero-shot transfer to unseen exchange-correlation functionals.

Without any SCF calculation, the predicted density recovers dipole and quadrupole moments to within a few percent of DFT references. This indicates that coefficient-space modeling improves not only the density itself but also directly supports estimation of observable response quantities. The abstract reports dipole and quadrupole moments within a few percent of DFT references, with no SCF performed.

The work targets settings where electron density is represented in an atom-centered GTO basis, and it applies to first-principles workflows that need scalable, transferable SCF initialization, such as density prediction on molecular datasets followed by self-consistent field solving. Its benefits are expressed in the abstract through density error on QM9 and the MD benchmark, SCF iteration counts, and recovery of dipole and quadrupole moments, so the most direct audience is quantum-chemistry and materials-computing workflows that use a comparable basis and care about SCF convergence cost. The zero-shot transfer to unseen exchange-correlation functionals points to use in screening scenarios where functionals change frequently.

The abstract does not state the specific sizes of the QM9 and MD benchmarks, the training and evaluation splits, or the definition of the density error metric, nor does it give hyperparameters or ablations for the two-phase trajectory curriculum; it also does not specify the initial guess and convergence criteria under which the 68% SCF iteration reduction was measured. Whether density error and moment accuracy hold at the same level under zero-shot transfer to unseen exchange-correlation functionals needs confirmation from the main text. Because only the abstract is available here, these details are open questions to verify rather than grounds for dismissal.