NetObs

Calculate observables from neural network–based VMC (NN-VMC).

Currently, NetObs has built-in suport for FermiNet (molecular systems) and DeepSolid (solid systems). We are planning to add support for more neural network VMC frameworks, and it's also very easy to write your own adaptor to make NetObs support your network!

The currently supported observables include energy, electron density, interatomic force, stress, etc. More observables are on the way, and it is also very easy to write your own!

Installation

Clone this repository and run an editable installation:

shell pip install -e .

Please don't forget to install the NN-VMC code and the corresponding version of JAX beforehand. NetObs is tested with google-deepmind/ferminet@c6f9db5+jax@0.4.38 and bytedance/DeepSolid+jax@0.2.26. For more details, you can refer to our CI config.

Command Line Example

NetObs comes with a friendly command-line app, making it easy for a quick try. An example command would be:

shell netobs @ferminet_vmc tests/data/H_atom.py @force:SWCT --with steps=20 --net-restore tests/data/H_atom.npz

A short explanation of the arguments:

• @ferminet_vmc specifies the Adaptor. The @ sign is a shortcut to access built-in Adaptors, Estimators, etc. @ferminet_vmc stands for the DEFAULT object in module netobs.adaptors.ferminet_vmc. Another example would be my_module:MyAdaptor.
• tests/data/H_atoms.py is a Python module or file that contains a get_config function. If you have a custom Adaptor, it can be anything your Adaptor recognizes.
• @force:SWCT specifies the Estimator. @force:SWCT means the SWCT class in the netobs.observables.force module.
• --with steps=20 specifies the options for NetObs.
• --net-restore ... tells the Adaptor where to restore the network checkpoint.

Another example with DeepSolid:

shell netobs @deepsolid_vmc tests/data/H_chain.py @energy --with steps=20 --net-restore tests/data/H_chain.npz

[!NOTE] If you want to use the SWCT force estimator with DeepSolid, it is highly recommended to use the tri feature.

Core Concepts

Adaptor

Adaptor is an abstract layer over the network that exposes a uniform API. For example, it tells NetObs how to restore from a checkpoint, evaluates the network and Hamiltonian, etc.

Currently, we have built-in support for FermiNet (molecules) and DeepSolid (solid systems). You can find their implementation in netobs/adaptors. It is also very easy to write your own adaptors!

Estimator

You can regard estimators as different ways to estimate a physical quantity given a set of Monte Carlo samples. You can have multiple estimators for an observable, but usually a basic one is enough.

In the code, an Estimator tells us what type of observable it targets at, the expression of the estimator given a set of Monte Carlo samples, how to combine the results from different steps (simply averaging or more actions are required), etc.

An Estimator can work only in the molecular case, or it can be implemented to support molecules and solids at the same time. Check out the SWCT Estimator in force.py for more.

You can find the implementation of built-in estimators in netobs/observables. And it is easy to create your own estimator!

Options

• steps (int, default to 10): Steps to run inference.
• mcmc_steps (int, default to 10): Steps to run MCMC.
• mcmc_burn_in (int, default to 100): Burn-in steps for MCMC.
• random_seed (int, default to current time): The random seed for the Monte Carlo simulation.
• log_interval (int, default to 10): Time interval in seconds between logs.
• save_interval (int, default to 600): Time interval in seconds between saves.
• estimator (dict): Options for the estimators.
• observable (dict): Options for the observable.

Integrate in Your Code

You need to pass your Adaptor, Estimator class, and evaluate options to netobs.evaluate.evaluate_observable. For example,

```python from netobs.helpers.importer import importmoduleor_file from netobs.observables.force import SWCT

cfg = importmoduleorfile("Hchain.py").getconfig() adaptor = DeepSolidVMCAdaptor(cfg, []) options = NetObsOptions(steps=20) digest, allvalues, state = evaluate_observable(adaptor, SWCT, options=options) ```

For more details, check out this test for how to start an evaluation, and check out this test for how to use different adaptors in your code.

Contributing

Contributions are welcomed and highly appreciated! We are open to new network Adaptors and new Estimators! For a detailed contribution guide, please see CONTRIBUTING.md.

Citation

If you are using NetObs in your works, please consider citing our papers.