Computational Workflows
Computational workflows encode a multistep simulation procedure (structure generation, job submission, error handling, restarts and analysis) as a graph of tasks that a workflow engine such as FireWorks or AiiDA executes with little human supervision. They make high-throughput density functional theory practical and reproducible, record provenance, and can embed decision rules that choose which calculations to run next.
Scripted workflows and checkpointing in ASE (2017)
ASE replaces input-file-driven runs with Python scripts in which a sequence of calculations is just a loop, observers are attached as callbacks, and results are stored with key-value metadata in the ase.db database (JSON, SQLite3 or PostgreSQL backends, queryable from the command line or a web browser) (Hjorth Larsen et al. 2017). A checkpointing module built on that database caches energy and force evaluations so a long script can be rerun after hitting a queue wall-time limit and resume where it stopped, which the authors present as a basis for robust automated screening workflows; the genetic algorithm module similarly evaluates candidates in separate queued jobs coordinated through the database. The authors note that utilities for tracking the status of many screening calculations were still missing. Summary: hjorthlarsen2017.
WhereWulff: an autonomous OER workflow for oxides (2023)
WhereWulff encodes the full oxygen evolution reaction modeling procedure for oxides in a FireWorks, pymatgen, atomate and Custodian stack, from bulk magnetic-ordering search and Pourbaix stability through slab enumeration, surface energies, Wulff-shape facet prioritization, surface Pourbaix termination selection and overpotential calculation (Sanspeur et al. 2023). Its decision steps cut the DFT count for two double perovskites from about 240 to about 132, and a ContinueOptimizeFW task with FireWorks detours handled about 180 wall-time resubmissions automatically while passing geometry, moments and wavefunctions between jobs. An end-to-end regression test on IrO2 with simulated VASP runs on every merge, and the same mechanism allows restarting a workflow mid-graph. Summary: sanspeur2023.
High-throughput surface enumeration and nuclearity classification (2024)
Sharma and co-workers chained AFLOW structure retrieval, convex-hull filtering, pymatgen slab enumeration (Miller indices below three, all terminations) and a graph-based nuclearity and shape analysis into a workflow parallelized with Dask, processing 29,002 intermetallic surfaces and releasing the catalog and code (github.com/ulissigroup/NuclearityCalculation) for reuse in screening for site geometry (Sharma et al. 2024). Summary: sharma2024.