Semantic workflows for atomistic simulations

Atomistic simulations of mechanical and thermodynamic properties are often assembled from ad hoc scripts, with information about methods, parameters, provenance, and analysis distributed across input files and post-processing steps. This practice replaces that fragmentation with modular, reusable workflows that capture scientific metadata during the simulation process. The workflows cover equation of state, elastic tensors, mechanical loading, thermal properties, free energies, defect formation energies, and nanoindentation. CMSO describes computational samples, ASMO represents simulation methods, and PROV-O captures provenance. atomRDF converts the resulting metadata into an RDF knowledge graph that can be queried using SPARQL. Implementations with pyiron and jobflow, together with examples using LAMMPS and Quantum ESPRESSO, show that the semantic layer is independent of a specific workflow manager or simulation backend. Machine learning interatomic potentials, including ACE and GRACE, can be used alongside conventional potentials within the same framework. The practice is positioned as infrastructure for AI in materials research, since machine readable provenance is what makes simulation output usable by agentic systems.

This work was developed within the NFDI-MatWerk consortium, a German national research data infrastructure initiative for materials science and engineering. NFDI-MatWerk supports the development of common semantic approaches for materials research data, and the ontologies used here were developed in this context. The collaboration involves four institutions: Forschungszentrum Jülich, TU Clausthal, Ruhr-Universität Bochum, and the Bundesanstalt für Materialforschung und -prüfung, combining expertise in knowledge engineering and atomistic simulation. Computing time was provided on the JURECA supercomputer at Forschungszentrum Jülich. The design separates the semantic representation from both the workflow engine and the simulation backend. An equivalent implementation in jobflow demonstrates portability across workflow systems, while a Quantum ESPRESSO example demonstrates its application beyond LAMMPS. Metadata generated by individual workflow nodes are collected in a lightweight Python dictionary with keys aligned to ontology concepts and serialised into RDF using atomRDF. A validation step is performed before the results are incorporated into the knowledge graph, allowing users to assess their physical relevance. This step is currently manual but can be partially automated for individual properties. Iron was selected as the demonstration system because of the availability of well established reference data and a large set of interatomic potentials.

The framework was demonstrated on iron, chosen as a well characterised reference material, across workflows for mechanical and thermodynamic properties. Bulk modulus and elastic constants were compared across twelve interatomic potentials, with each result linked in the knowledge graph to the potential that produced it and its bibliographic reference. This enables systematic comparison between models while retaining the provenance of each result. Semantic annotation also makes information accessible that is often implicit in conventional simulation workflows: because a substitutional defect is annotated when it is introduced, a single query can retrieve every sample containing one without reconstructing this information from atomic coordinates. Querying grain size against flow stress across the stored nanocrystalline iron simulations reproduced the Hall–Petch and inverse Hall–Petch regimes. The workflows are published on GitHub, an interactive Binder instance allows them to be explored without local installation, and the associated data and metadata are deposited on Zenodo with a persistent DOI. By retaining simulation results together with their provenance and semantic context, the framework produces machine readable data that can be reused in AI based analysis and agentic materials research.

Azócar Guzmán, A., Luu, H.T., Menon, S., Hickel, T., Merkert, N. and Sandfeld, S., 2026. Toward Knowledge‐Based Workflows: A Semantic Approach to Atomistic Simulations for Mechanical and Thermodynamic Properties. Advanced engineering materials, p.e70869., https://doi.org/10.1002/adem.70869