Scalable quantum simulations of molecular systems via improved optimization of neural quantum states

Quantum simulations of molecular systems hold transformative potential for computational chemistry, yet optimization inefficiencies and classical computational bottlenecks hinder practical implementation. We present algorithmic enhancements to the optimization of the unitary-coupled restricted Boltzmann machine Ansatz in the context of quantum machine learning, integrating adaptive learning rate and block optimization with the variational quantum imaginary time evolution algorithm. These improvements address convergence robustness and classical overhead in hybrid quantum-classical workflows. Demonstrations on small molecular systems show that our adaptive learning rate approach achieves chemically accurate results with fewer optimization steps compared to conventional methods, while block optimization further enables efficient parameter updates for larger systems, alleviating classical bottlenecks without compromising quantum expressivity. These advancements offer the possibility of extending the reach of near-term quantum hardware to scalable molecular simulations.