QMCTorch

• Easily use and implement new neural network ansatz
• Use ADF or pySCF as SCF backend
• Use Horovod to deploy on GPU clusters

In QMCTorch the trial wave function is calculated by a small, physically motivated network. Starting from the electronic positions, R, the first layer computes the values of all atomic orbitals for all electrons. From there a linear map computes the values of all relevant molecular orbitals. A Slater Pooling mask is then applied to compute all Slater determinants that are finally combined by a fully connected layer. The Jastrow factors are computed in parallel and combined with the CI expansion to obtain the value of the wave function