Skip to content

FermiNet Tutorial: Calculating the Ground State Energy of the H2 Molecule Using DeepChem

In this blog, we will demonstrate how to use the FermiNet model implemented in Deep Chemistry (DeepChem) to calculate the ground state energy of the hydrogen (H2) molecule. We will cover the following steps:

  • Launch the Instance in Alces Cloud and install Jupyter Notebook.
  • Installing required packages
  • Initializing and pretraining the FermiNet model
  • Training the FermiNet model to obtain the ground state energy

Launch the Instance

All the steps to launch and connection to instance is provided in link.

Install Jupyter Notebook and Lab

All the steps to install jupyter notebook and lab is provided in link.

Installing Required Packages

First, make sure you have installed the latest version of Python and pip. Next, run the following commands to install the necessary packages: 

!pip install --pre deepchem
!pip install torch_geometric
!pip install pyscf

Initializing and Pretraining the FermiNet Model

To initialize the FermiNet model, provide a list of lists containing the nuclear coordinates in the following format: [[element symbol, [3D coordinates]]]. For example, to define an H2 molecule, create the variable H2_molecule as follows: 

H2_molecule = [['H', [0, 0, 0]], ['H', [0, 0, 1.4135151]]]

Next, initialize the FermiNet model with the desired configuration. In our case, we set the spin to 0, ionic charge to 0, and four batches to improve accuracy.

from deepchem.models.torch_models.ferminet import FerminetModel
import torch
import logging

logger = logging.getLogger()
logger.setLevel(logging.DEBUG)

mol = FerminetModel(H2_molecule, spin=0, ion_charge=0, batch_no=4)
mol.train(nb_epoch=150)
# After pretraining, verify the obtained energy:

mol.model.forward(torch.from_numpy(mol.molecule.x))
energy = mol.model.calculate_electron_electron() \
          - mol.model.calculate_electron_nuclear() \
          + mol.model.nuclear_nuclear_potential \
          + mol.model.calculate_kinetic_energy()
mean_energy = torch.mean(energy)
print("Energy before start:", mean_energy)

This should produce an energy close to the Hartrees-Fock (HF) ground state energy.

Training the FermiNet Model to Obtain the Ground State Energy

Before beginning the actual training process, ensure that the "prepare_train" method has been called to perform Markov Chain Monte Carlo (MCMC) burn-in and reinitialize electron positions. Then, call the "train" method to initiate the FermiNet training.

import logging

logger = logging.getLogger()
logger.setLevel(logging.DEBUG)

mol.prepare_train()
mol.train(nb_epoch=100, lr=0.0002, std=0.04)

Once completed, access the final_energy attribute of the mol object to view the net average ground state energy over all training epochs.

mol.final_energy.item()

This value should be close to the expected theoretical value of approximately 1.174476 Hartrees. To achieve higher precision, increase the number of iterations or fine-tune other hyperparameters such as MCMC step proposals and layer sizes.