ABSTRACT:

Neural Network Quantum States (NNQS) are a recent and powerful development in computational chemistry. They represent the electronic wavefunction as a neural network optimized via Variational Monte Carlo (VMC), providing a flexible and systematically improvable ansatz capable of achieving near-exact energies for strongly correlated molecules. Recent foundation-model approaches have shown that a single pre-trained NNQS can generalize across molecular Hamiltonians and fine-tune to chemical accuracy 6–16× faster than training from scratch. Despite this progress, fundamental questions remain: how does NNQS accuracy scale with architecture and compute across correlation regimes; when is foundation-model fine-tuning preferable to single-system transfer learning; and how far can current models be extended to chemically distinct domains?

This proposal addresses these questions through three integrated directions motivated by a need for accurate modelling of chemical processes involved in CO₂ reduction and N₂ fixation—industrially crucial processes where strong electronic correlation limits existing methods and reliable reference data is scarce.

First, this proposal will establish empirical scaling laws for NNQS accuracy as a function of architecture, parameter count, and computational effort across weakly to strongly correlated systems, and develop principled learning-rate transfer rules to reduce costly hyperparameter retuning across model scales. Second, it will benchmark foundation-model fine-tuning against single-system transfer learning along reaction coordinates, providing quantitative guidance on convergence speed and final accuracy in realistic chemical workflows. Third, it will extend the Orbformer foundation model to explicitly include the ionic and radical intermediates central to green chemistry.

Expected outcomes are: (1) validated scaling laws for NNQS across correlation regimes; (2) a practical transfer-learning protocol for molecular systems; (3) an extended Orbformer checkpoint for green chemistry targets; and (4) near-exact wavefunction reference data for key intermediates in CO₂ reduction and N₂ fixation.