Ontology-based knowledge graph infrastructure for interoperable atomistic simulation data

This paper presents an ontology-based knowledge graph infrastructure that integrates heterogeneous atomistic simulation data and workflows into a standardized, machine-readable format to enhance data findability, interoperability, and reuse.

The Gist

Imagine you are a chef trying to recreate a famous dish from a recipe book. But here's the catch: every time you look at a different book, the ingredients are listed in different languages, the measurements are in random units (cups, grams, "a pinch"), and the instructions are written in code only the original chef understands. Worse yet, the books don't tell you why the chef chose those specific ingredients or what happened if they made a mistake.

This is exactly the problem scientists face with atomistic simulation data. These are computer experiments that simulate how atoms behave to create new materials. Currently, this data is scattered across thousands of files, stored in incompatible formats, and often missing the "story" of how it was made. It's like having a library where every book is written in a different dialect, and half the pages are torn out.

This paper presents a solution: A Universal Translator and a Master Librarian for Atomic Data.

Here is how they did it, explained simply:

1. The "Universal Dictionary" (The Ontologies)

First, the team built a massive, shared dictionary called an Ontology. Think of this as a strict set of rules for how to describe things.

• Before: One scientist might call a material "Iron-Defect-Alpha," while another calls it "Fe-Vacancy-01." A computer has no idea these are the same thing.
• After: The Ontology says, "No matter what you call it, if it's Iron with a missing atom, we will all call it Iron:Vacancy."
• They created two main dictionaries:
  • CMSO: Describes the "ingredients" (the materials, the atoms, the defects).
  • ASMO: Describes the "cooking method" (the software used, the math formulas, the steps taken).

2. The "Smart Translator" (The Software Infrastructure)

Even with a dictionary, scientists don't want to stop using their familiar tools (like Excel or Python scripts) to write their notes. They don't want to learn a new, complex language just to be "correct."

So, the team built a middleware layer (called atomRDF).

• Imagine a translator sitting between the scientist and the library.
• The scientist writes their notes in their usual format (YAML, JSON, or even a simple text file).
• The translator instantly converts those notes into the "Universal Dictionary" format and files them away in a giant, connected database called a Knowledge Graph.
• The scientist doesn't have to change their workflow; the system does the heavy lifting in the background.

3. The "Giant Connected Web" (The Knowledge Graph)

Instead of storing data in separate, isolated folders, they built a Knowledge Graph.

• Old Way: Data is like a stack of index cards in different drawers. To find a connection, you have to manually pull out card A, then card B, and hope they match.
• New Way: The Knowledge Graph is like a giant, glowing spiderweb. Every piece of data (an atom, a temperature, a software code) is a node. Every relationship (was calculated by, is made of, depends on) is a string connecting them.
• Because everything is connected, you can ask the web complex questions like: "Show me all the energy calculations for Copper defects made using Method X, but only if the temperature was above 500 degrees." The web lights up and gives you the answer instantly, even if that data came from five different research groups.

What Can You Do With This? (The Magic Tricks)

The paper shows three cool things this system can do:

• The "Detective" (Cross-Dataset Analysis):
  They took data about "grain boundaries" (where crystals meet) from many different sources. Because the data was standardized, they could instantly see patterns that were invisible before. For example, they could see that certain types of boundaries are stable in Copper but not in Aluminum, simply by querying the web. It's like being able to compare every recipe for "Chocolate Cake" ever written to find the perfect one instantly.

• The "Time Traveler" (Deriving New Science):
  Sometimes scientists calculate data but forget to calculate the final result. The team found old data about how materials expand when heated. By connecting the dots in the graph (Volume + Temperature + Time), they mathematically derived a new property (Thermal Expansion) that the original authors never explicitly published. They turned "dust" into "gold."

• The "Replay Button" (Provenance & Reconstruction):
  This is perhaps the most powerful feature. In science, knowing how you got a result is as important as the result itself.

  • The system records the entire "cooking video" of the simulation.
  • If you find a result, you can press "Rewind" and see exactly which software, which version, and which settings were used.
  • Better yet, the system can try to rebuild the recipe automatically. It can generate a new script that says, "Here is the code to recreate this exact experiment." This solves the "it worked on my computer" problem forever.

Why Does This Matter?

Currently, a lot of scientific data is "orphaned"—it exists, but it's too messy to use again. This infrastructure turns that messy pile of data into a FAIR resource:

• Findable (You can search for it easily).
• Accessible (Anyone can get it).
• Interoperable (Different computers can talk to each other).
• Reusable (You can use old data to do new science).

In short, this paper builds the operating system for the future of materials science. It stops scientists from wasting time translating files and starts them spending time discovering new materials, from better batteries to stronger metals, by letting their computers do the organizing.

Technical Summary

1. Problem Statement

The reuse and integration of atomistic simulation data in materials science are currently hindered by several critical bottlenecks:

• Heterogeneity: Data is stored in software-specific formats (e.g., VASP, LAMMPS, QuantumEspresso) with inconsistent metadata, making cross-platform interoperability difficult.
• Lack of Standardization: Workflow descriptions, provenance (computational history), and simulation parameters are often incomplete, implicit, or unstructured.
• Complexity of Defect Systems: While databases exist for bulk materials (e.g., Materials Project), they struggle with defect-containing systems (grain boundaries, vacancies) where descriptions depend heavily on local atomic environments and specific simulation workflows.
• Reproducibility Gap: Reconstructing a calculation from published results often requires substantial manual effort to cross-reference input files, scripts, and publication text, breaking the chain of reproducibility.

2. Methodology

The authors propose a modular, ontology-based infrastructure that bridges the gap between raw simulation data and machine-interpretable knowledge graphs. The methodology consists of three main layers:

A. Ontology Framework

Two primary ontologies were developed to provide a shared semantic schema:

1. Computational Materials Sample Ontology (CMSO): Describes material structures from the atomic to the macroscale. It covers crystallography, chemistry, defects, and simulation cells. It is modular, allowing extensions for different length scales (Nano, Meso, Micro, Macro).
2. Atomistic Simulation Methods Ontology (ASMO): Describes computational methods (DFT, Molecular Dynamics, Kinetic Monte Carlo), workflows, and provenance. It builds on the W3C PROV-O standard to track activities, agents, and entities. It also integrates QUDT for unit handling and MDO for electronic structure concepts.

B. Software Infrastructure (The "Stack")

To make ontologies usable in routine scientific workflows without requiring direct interaction with complex RDF/OWL syntax, the authors developed a layered software pipeline:

1. Conceptual Metadata Capture (conceptual_dictionary): A lightweight, ontology-aligned layer using YAML/JSON templates and Python dictionaries. This allows scientists to capture metadata in familiar formats (human- and machine-readable) without needing RDF expertise. It supports manual entry and automated parsing from legacy files.
2. Ontology-Aligned Data Models (atomRDF): A translation layer using Pydantic data classes. This layer validates the metadata from the conceptual layer, ensuring type safety and consistency. It includes bidirectional methods (to_graph and from_graph) to serialize data into RDF triples or reconstruct Python objects from a graph.
3. Knowledge Graph Construction: Data is serialized into an RDF graph (using rdflib) where entities are linked via persistent identifiers (IRIs/UUIDs).

C. FAIR Alignment

The system is designed to satisfy FAIR principles (Findable, Accessible, Interoperable, Reusable) by using globally unique identifiers, SPARQL endpoints, open licenses, and explicit provenance modeling.

3. Key Contributions

• Unified Semantic Representation: A framework that normalizes heterogeneous atomistic data (structures, workflows, properties) into a single, queryable knowledge graph.
• Two-Way Provenance: The system supports forward tracking (capturing provenance at the point of data generation) and backward reconstruction (reconstructing computational workflows from existing results).
• Practical Adoption Strategy: By introducing an intermediate "conceptual dictionary" layer, the authors lower the barrier to entry for domain scientists, allowing integration with existing tools (e.g., AiiDA, pyiron) without forcing immediate migration to raw RDF.
• Defect-Centric Focus: Unlike many existing databases focused on bulk materials, this infrastructure explicitly models crystallographic defects and their complex simulation workflows.

4. Results and Demonstrations

The authors integrated data from multiple sources (Zenodo, GitHub, publications) and generated new data to demonstrate the system's capabilities:

• Scale: The resulting knowledge graph contains 757,253 triples describing 7,926 computational samples.
• Semantic Integration (Grain Boundaries): The system successfully integrated grain boundary data from disparate sources. It enabled targeted SPARQL queries (e.g., "Find all $\Sigma3$ grain boundary energies calculated via DFT") that revealed data gaps and trends impossible to see in isolated datasets.
• Cross-Dataset Analysis: The graph allowed for the correlation of unrelated properties. For example, the authors identified a positive correlation between vacancy formation energy and grain boundary energy across different elements and boundary types.
• Derivation of New Properties: By querying NPT ensemble MD simulations, the authors extracted volumetric thermal expansion coefficients for various elements (Si, Li, Al, Fe, Ge) from data that originally only reported volume and temperature, demonstrating the ability to generate new scientific value from existing data.
• Workflow Reconstruction: The system successfully reconstructed a molecular statics workflow for vacancy formation energy. While it highlighted missing links (e.g., specific interatomic potential file versions), it proved that the logic of the workflow (scaling energies, calculating differences) could be automatically inferred and re-executed.

5. Significance and Impact

• Interoperability: The work solves the "silo" problem in computational materials science, allowing data from different codes and formats to be compared and analyzed jointly.
• Reproducibility: By capturing the full computational history (including post-processing steps often done manually in scripts), the infrastructure moves the field toward computational reproducibility, not just data availability.
• Foundation for AI/ML: The structured, machine-readable nature of the knowledge graph provides high-quality training data for machine learning models, particularly for discovering materials with defects.
• Scalability: The modular design allows the ontology to evolve with new scientific requirements (e.g., new simulation methods or defect types) without breaking the existing infrastructure.

Limitations & Future Work:
The authors note that the system currently relies on the quality of input metadata; legacy data with missing provenance requires manual curation. Future work aims to automate metadata extraction using Large Language Models (LLMs), improve the representation of external dependencies (like interatomic potential files), and expand the ontology to cover a broader range of multiscale phenomena.