Abstract:Finding accurate solutions to the electronic Schrödinger equation plays an important role in discovering important molecular and material energies and characteristics. Consequently, solving systems with large numbers of electrons has become increasingly important. Variational Monte Carlo (VMC) methods, especially those approximated through deep neural networks, are promising in this regard. In this paper, we aim to integrate one such model called the FermiNet, a post-Hartree-Fock (HF) Deep Neural Network (DNN) model, into a standard and widely used open source library, DeepChem. We also propose novel initialization techniques to overcome the difficulties associated with the assignment of excess or lack of electrons for ions.

What problem does this paper attempt to address?

The paper mainly discusses how to integrate FermiNet, a post-Hartree-Fock (HF) deep neural network (DNN) model, into the widely used open-source library DeepChem to solve the exact solution problem of the electronic Schrödinger equation, especially for solving a large number of electronic systems. In quantum mechanics, finding the expected energy of the trial wave function is always greater than or equal to the ground state energy of the molecular system, and variational Monte Carlo (VMC) methods, especially those based on deep learning, have great prospects for this. The paper proposes a new initialization technique to address the problem of excess or missing electron allocation in ions. Traditional VMC methods rely on Monte Carlo algorithm initialization and sampling of electronic coordinates, and then optimize the expected energy step by step. The authors also introduce DeepChem, an open-source Python library for scientific machine learning and deep learning in molecular and quantum data, which supports various physical calculations and has integrated different differentiable density functional theory infrastructure. In the implementation part, the paper describes in detail the components of the FermiNet model, including the electronic sampler, HF baseline, ion charge initialization, and FermiNet model (neural network layers). In particular, ion charge initialization is proposed through Mulliken charge analysis to handle the number of electrons in ions more accurately. The training of the FermiNet model consists of supervised pre-training and unsupervised training. In the pre-training phase, orbital values calculated using the HF method are used as labels, and in the unsupervised training phase, the model parameters are adjusted based on the expected energy. The experimental results show that the ground state energy curve of the FermiNet model closely matches the CCSD (coupled cluster) method and outperforms the HF method in the calculation of the ionization potential of LiH molecules, approaching the calculation values of CCSD. The conclusion of the paper points out that integrating FermiNet into DeepChem can promote fast experiments and improve the ability to accurately calculate the ground state energy of molecular/ion systems. Future work plans to improve the model using PyTorch's JIT (Just-in-Time) functionality and combined with the characteristics of PauliNet to achieve faster convergence speed and accuracy.

Open-Source Fermionic Neural Networks with Ionic Charge Initialization

Fermionic neural network with effective core potential

Unsupervised Deep Neural Network Approach To Solve Fermionic Systems

Ab initio quantum chemistry with neural-network wavefunctions

CHGNet: Pretrained universal neural network potential for charge-informed atomistic modeling

High-Dimensional Neural Network Potential for Liquid Electrolyte Simulations.

Efficient solutions of fermionic systems using artificial neural networks

Towards a transferable fermionic neural wavefunction for molecules

CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling

Better, Faster Fermionic Neural Networks

A unified quantum framework for electrons and ions: The self-consistent harmonic approximation on a neural network curved manifold

Interatomic Force from Neural Network Based Variational Quantum Monte Carlo

Scalable molecular simulation of electrolyte solutions with quantum chemical accuracy

Towards a Foundation Model for Neural Network Wavefunctions

SchrödingerNet: A Universal Neural Network Solver for The Schrödinger Equation

Ju l 2 01 8 Solving Many-Electron Schrödinger Equation Using Deep Neural Networks

Neural network variational Monte Carlo for positronic chemistry

Including physics-informed atomization constraints in neural networks for reactive chemistry

Open Source Infrastructure for Differentiable Density Functional Theory

Charge-Optimized Electrostatic Interaction Atom-Centered Neural Network Algorithm