Ab initio foundation models for transferable deep quantum Monte Carlo simulations of molecular wave functions.

Computing accurate yet efficient approximations to the many-electron Schrödinger equation lies at the core of computational chemistry.

Variational Monte Carlo (VMC) is a versatile ab initio method to simulate ground and low-lying excited states, using explicitly correlated wave functions and stochastic sampling.

Neural network parametrizations of molecular trial wave functions have recently elevated the accuracy of VMC to rank among the most precise electronic structure methods, albeit with high computational cost per system.

Leveraging the high expressivity of neural network wave functions, transferable VMC introduces a paradigm shift relative to traditional quantum chemistry methods by exploiting molecular similarities to amortize cost across systems.

We present Orbformer, a chemically transferable neural network wave function trained on 22,000 equilibrium and dissociating molecular structures.

Orbformer is a “foundation model” for small molecules that can be fine-tuned on unseen compounds at a fraction of the cost of training from scratch.

It delivers highly accurate, systematically improvable relative energies on challenging benchmarks and achieves accuracy–cost ratios competitive with classical methods in multi-reference regimes.

By leveraging large-scale pretraining, we demonstrate for the first time that deep-learning-based quantum Monte Carlo is a practical and cost-effective tool in quantum chemistry.