Finding accurate solutions to the Schrödinger equation is one of the key unsolved challenges of computational chemistry. Recently combining variational Monte Carlo with a deep-learning ansatz has emerged as a promising way to obtain highly accurate energies at moderate scaling of computational cost. We present a novel architecture for fermionic wavefunctions, which achieves substantially higher accuracy at 5-10x lower computational cost compared to previous approaches. Using our method we calculate the most accurate variational ground state energies ever published for a number of different atoms and small molecules. We systematically break down our improvements, focusing in particular on the effect of physical prior knowledge, where we surprisingly find that increasing the prior knowledge given to the architecture can actually decrease accuracy.
When calculating solutions for multiple molecular geometries (e.g. to study a Potential Energy Surface) we can furthermore accelerate VMC-optimization. We demonstrate how sharing weights across wavefunction models during optimization, reduces the computational cost by an order of magnitude.
Back to Workshop IV: Monte Carlo and Machine Learning Approaches in Quantum Mechanics