WorkFlow for Automated Cluster Expansion Regression (WFacer)

Modulated automation of cluster expansion model fitting based on atomate2 and Jobflow

---

WFacer ("Wall"Facer) is a light-weight package based on smol to automate the fitting of lattice models in disordered crystalline solids using cluster expansion method. Beyond metallic alloys, WFacer is also designed to handle ionic systems through enabling charge Decorator and the external EwaldTerm. Powered by Atomate2, Jobflow and Fireworks, WFacer is able to fully automate the cluster expansion building process on super-computing clusters, and can easily interface with MongoDB data storage in the Materials Project style .

Functionality

WFacer currently supports the following functionalities:

• Preprocess setup to a cluster expansion workflow as dictionary.
• Enumerating and choosing the least aliasing super-cell matrices with given number of sites; enumerating charge balanced compositions in super-cells; Enumerating and selecting low-energy, non-duplicated structures into the training set at the beginning of each iteration.
• Computing enumerated structures using Atomate2 VASP interfaces.
• Extracting and saving relaxed structures information and energy in Atomate2 schemas.
• Decorating structures. Currently, supports charge decoration from fixed labels, from Pymatgen guesses, or from a gaussian optimization process based on partitioning site magnetic moments.
• Fitting effective cluster interactions (ECIs) from structures and energies with sparse linear regularization methods and model selection methods provided by sparse-lm, except for overlapped group Lasso regularization.
• Checking convergence of cluster expansion model using the minimum energy convergence per composition, the cross validation error, and the difference of ECIs (if needed).
• Creating an atomate2 style workflow to be executed locally or with Fireworks.

Installation

• From pypi: pip install WFacer
• From source: Clone the repository. The latest tag in the main branch is the stable version of the code. The main branch has the newest tested features, but may have more lingering bugs. From the top level directory, do pip install -r requirements.txt, then pip install . If you wish to use Fireworks as the workflows manager, do pip install -r requirements-optional.txt as well.

Post-installation configuration

Specific configurations are required before you can properly use WFacer.

• We highly recommend to use Fireworks but it is not required. To use job management with Fireworks and Atomate2, configuring Fireworks and Atomate2 with your MongoDB storage is necessary. Users are advised to follow the guidance in Atomate2 and Atomate installation guides, and run a simple test workflow to see if it is able to run on your queue.

  Instead of writing in my_qadapter.yaml as

     rlaunch -c <<INSTALL_DIR>>/config rapidfire
  

  we suggest using:

     rlaunch -c <<INSTALL_DIR>>/config singleshot
  

  because by using singleshot with rlaunch, a task in the submission queue will be terminated once a structure is finished instead of trying to fetch another structure from the launchpad. This can be used in combination with:

     qlaunch rapidfire -m <number of tasks to keep in queue>
  

  to guarantee that each structure is able to use up the maximum wall-time in its computation.

• A mixed integer programming (MIP) solver would be necessary when a MIQP based regularization method is used. A list of available MIP solvers can be found in cvxpy documentations. Commercial solvers such as Gurobi and CPLEX are typically pre-compiled but require specific licenses to run on a super-computing system. For open-source solvers, the users are recommended to install SCIP in a dedicated conda environment following the installation instructions in PySCIPOpt.

A quick example for fully automating cluster expansion

A simple workflow to run automatic cluster expansion in a Ag-Li alloy on FCC lattice is as follows (see other available options in the documentations of preprocessing.py.):

from fireworks import LaunchPad
from WFacer.maker import AutoClusterExpansionMaker
from jobflow.managers.fireworks import flow_to_workflow
from pymatgen.core import Structure

# construct a rock salt Ag-Li structure
agli_prim = Structure(
    lattice=[[0, 2.13, 2.13], [2.13, 0, 2.13], [2.13, 2.13, 0]],
    species=[{"Ag": 0.5, "Li": 0.5},
             {"Ag": 0.5, "Li": 0.5},],
    coords=[[0, 0, 0], [0.5, 0.5, 0.5]],
)
# Use default for every option.
ce_flow = AutoClusterExpansionMaker(name="agli_fcc_ce", options={}).make(agli_prim)

# convert the flow to a fireworks WorkFlow object
# If argument "store" is not specified, all documents will be saved to the JOB_STORE
# Defined by the local configuration files where you run THIS script from.
wf = flow_to_workflow(ce_flow)

# submit the workflow to the FireWorks launchpad
lpad = LaunchPad.auto_load()
lpad.add_wf(wf)

After running this script, a workflow with the name "agli_fcc_ce" should have been added to Fireworks' launchpad.

Submit the workflow to queue using the following command once you have correctly configured Fireworks queue adapter,

nohup qlaunch rapidfire -m {n_jobs} --sleep {time} > qlaunch.log

where n_jobs is the number of jobs you want to keep in queue, and time is the amount of sleep time between two queue submission attempts. qlaunch will keep submitting jobs to the queue until no unfinished job could be found on launchpad.

Note: You may still need to qlaunch manually after every cluster expansion iteration because for Fireworks could occasionally set the enumeration job to the READY state but fails to continue executing the job.

After the workflow is finished, use the following codes to retrieve the computed results from MongoDB:

from jobflow import SETTINGS
from pydantic import parse_obj_as
from WFacer.schema import CeOutputsDocument

store = SETTINGS.JOB_STORE
store.connect()

# Just a random example. You have to check what is your maximum iteration on your own.
max_iter = 10
# Find the output of a trigger job, which should be the CeOutputDocument of the final
# iteration.
job_return = store.query_one({"name": f"agli_fcc_ce_iter_{max_iter}_trigger"})
raw_doc = job_return["output"]
# De-serialize everything.
doc = parse_obj_as(CeOutputsDocument, raw_doc)

# See WFacer.schema for more.
print("Cluster subspace:", doc.cluster_subspace)
print("Wrangler:", doc.data_wrangler)
print("coefficients:", doc.coefs_history[-1])

Note: Check that the Jobflow installations on the computer cluster and the query terminal are configured to use the same JOB_STORE.

Copyright Notice

Workflow for automated cluster expansion regression (WFacer) Copyright (c) 2023, The Regents of the University of California, through Lawrence Berkeley National Laboratory (subject to receipt of any required approvals from the U.S. Dept. of Energy) and the University of California, Berkeley. All rights reserved.

If you have questions about your rights to use or distribute this software, please contact Berkeley Lab's Intellectual Property Office at [email protected].

NOTICE: This Software was developed under funding from the U.S. Department of Energy and the U.S. Government consequently retains certain rights. As such, the U.S. Government has been granted for itself and others acting on its behalf a paid-up, nonexclusive, irrevocable, worldwide license in the Software to reproduce, distribute copies to the public, prepare derivative works, and perform publicly and display publicly, and to permit others to do so.