deepqmc/deepqmc

What it solves

DeepQMC provides a way to simulate electronic ground and excited states of molecules using deep neural network trial wave functions. It solves the molecular Hamiltonian, allowing for the variational optimization of these wave functions to accurately model molecular behavior.

How it works

Built on JAX and Haiku, the software uses deep learning ansatze to represent wave functions. It employs a penalty-based approach to optimize for excited states and a spin penalty to target specific spin sectors. It also supports geometric transferability, meaning a single neural network can be trained across multiple molecular configurations, either through fixed datasets or dynamic sampling.

Who it’s for

It is designed for researchers and scientists in computational chemistry and quantum chemistry who need to perform variational Monte Carlo simulations of molecular electronic states.

Highlights

• Supports a wide range of neural network wave function ansatze, including FermiNet, PauliNet, Psiformer, DeepErwin, and LapNet.
• Implements geometric transferability for optimization across multiple atomic configurations.
• Capable of simulating both ground and excited electronic states.
• Integrated with JAX for high-performance computing on CPUs and GPUs.