Speaker
Description
Computational materials science workflows often involve numerous steps, integrate diverse software tools, span multiple length and time scales, and cover a broad range of material compositions, structures, and thermodynamic conditions. Therefore, ensuring reproducibility, data reusability, and meaningful interpretation requires not only detailed descriptions of data and metadata at each stage of the workflow but also complete provenance. This includes precise documentation of software versions, computational environments, and the execution order of workflow steps. These requirements motivate the development of FAIR computational workflows, which extend the FAIR data and FAIR4RS principles to address the specific needs of computational reproducibility and reusability.
We present design concepts for a set of automated, reproducible, and FAIR workflows implemented within the pyiron workflow environment. Pyiron provides both a graphical user interface and a Jupyter notebook-based interface, lowering the barrier for composing and managing workflows and facilitating the integration of diverse software tools. Workflow outputs are annotated using open formats, and provenance information is recorded throughout. Each step functions as a self-contained computational node with version control and automatic resolution of software dependencies. These nodes are enriched with metadata, including references to the underlying computational methods.
As an example, we present a set of workflows for computing phase diagrams with near ab initio accuracy, employing machine learning interatomic potentials. These automated workflows include all steps from the generation of ab initio reference datasets to the parameterisation of the interatomic potentials, calculation of the phase diagram, and comparison with the CALPHAD approach. This example demonstrates how complex simulation protocols can be combined in a reproducible manner to create workflows that follow the FAIR principles.