Optimizing Neural Network Backflow for Scalable Ab-Initio Quantum Chemistry

Neural quantum states have recently advanced rapidly, emerging as powerful tools in computational quantum many-body physics. However, their application to quantum chemistry remains challenging, largely due to scalability issues such as the quartic growth of terms in second-quantized Hamiltonians, which restricts their use in larger molecular systems. In this work, we optimize neural network backflow (NNBF) for ab-initio quantum chemistry by developing scalable methods to reduce computational costs, benchmarking these enhancements on large molecules and strongly correlated systems. Our strategies include simplifying local energy computations, utilizing symmetries to improve the ansatz's expressiveness, and adopting more effective optimization techniques. These developments enable more efficient ab-initio simulations of large molecules.