Reusable Atomistic Workflows

• Reusable atomistic workflows are computational frameworks that organize atomistic simulations into modular, versioned nodes for clear provenance and reusability.
• They use declarative configurations and standardized I/O schemas to enable interoperability between diverse simulation engines like DFT, MD, and machine learning potentials.
• These workflows drive high-throughput materials studies by integrating FAIR data principles, semantic metadata, and automated resource management.

Reusable atomistic workflows are computational frameworks and methodologies that structure, parameterize, and document atomistic simulations in a way that enables transparent reuse, modular extension, full provenance, and reliable interoperability across codes, methods, and research teams. Central to this paradigm are common abstractions: workflow nodes as functional computational units, versioned interfaces for I/O, formal data provenance, and semantic metadata that together support the definition, execution, and sharing of automated atomistic pipelines across density functional theory (DFT), machine learning potentials (MLP), molecular dynamics (MD), and high-throughput property calculation.

1. Core Architectural Principles and Workflow Abstractions

Reusable atomistic workflows are defined as directed acyclic graphs (DAGs) whose nodes represent atomic-scale tasks (e.g., geometry optimization, energy calculation, property extraction), with clearly specified input and output schemas. This formalism is implemented in major platforms including ASE/ASR (Gjerding et al., 2021), AiiDA (Huber et al., 2021), ChemGraph (Pham et al., 3 Jun 2025), Sim2Ls (Hunt et al., 2021), and QUASAR (Yang et al., 30 Jan 2026). Workflows are typically parameterized via high-level automation interfaces (Python scripts, YAML/JSON specs, or visual editors), supporting composition of simulation subtasks, dynamic dependency resolution, and scalable orchestration.

A canonical schema is illustrated by QUASAR's representation $W = (N, E)$, where each node $n_i$ has associated input/output definitions, versioning, computational cost, and provenance pointers. Platforms such as ASR encapsulate simulation steps as "Recipes," defined by decorated Python functions with automatic wrapping into versioned, cacheable, and composable workflow nodes (Gjerding et al., 2021).

Key abstractions across frameworks include:

• Node abstraction: Encapsulates a simulation step as an independently versioned, modular callable object, supporting parameterization and I/O schema validation (ASR Instruction, MiqroForge node.json, AiiDA WorkChain).
• Data provenance: Comprehensive recording of all inputs, code versions, intermediate and final outputs with unique identifiers (UUIDs), typically stored in project-local or global databases.
• Declarative configuration: Input/output schemas are explicitly enumerated, validated for type, shape, units, and default values, often in machine-readable formats (YAML/JSON, INPUTS cells in Sim2Ls).
• Engine abstraction: Swappable interfaces that decouple workflow logic from underlying simulation engines (ASE Calculators, AiiDA engine adapters, MiqroForge Docker nodes).
• Cache and record management: Hierarchical or database-backed caches prevent recomputation and enable rapid rerun or resumption with full reproducibility.

2. Provenance, Versioning, and FAIR Principles

Reusable atomistic workflows enforce FAIR (Findable, Accessible, Interoperable, Reusable) data principles by:

• Automatic provenance capture: Every action, parameter, code version, and result is recorded with cryptographic hashes and timestamps (ASR Record, Sim2L ResultsDB, AiiDA provenance graph, ChemGraph message state).
• Semantic, ontology-aligned metadata: Annotations follow community or application ontologies (ASMO, CMSO, PROV-O) linking each data instance to simulation methods, materials, parameters, and agents, as in the knowledge-based workflow framework (Guzman et al., 1 Feb 2026).
• Result and workflow discovery: Published and versioned workflows are indexed in global platforms (e.g., nanoHUB for Sim2Ls, workflow libraries in ChemGraph, registry in MiqroForge), with metadata including DOI, license, authorship.
• Data reproducibility guarantees: Sim2Ls, ASR, and MiqroForge snapshot all dependencies, tool/container versions, and parameterizations at runtime, enabling bitwise-reproducible reruns and transparent validation of prior results (Hunt et al., 2021, Wang et al., 11 Aug 2025).

3. Interoperability and Extensibility: Code-Agnostic Workflows

A hallmark of reusability is engine-agnostic execution. This is achieved by:

• Universal schema and adapters: For instance, the interoperability framework for DFT workflows prescribes a universal JSON/YAML schema for all input/output, and code-specific adapters translate this spec to and from CASTEP, GPAW, Quantum ESPRESSO, VASP, and other engines (Steensen et al., 14 Nov 2025).
• Common WorkChain architectures: In AiiDA, the CommonRelaxWorkChain exposes all workflow-level and engine-specific parameters, while the higher-level property workflows (equation of state, dissociation curve) reuse the relax task via standardized interfaces (Huber et al., 2021).
• Flexible chaining and modular composition: Modular design enables arbitrary chaining (e.g., HopDec's sequential modules for structure generation, defect creation, MD sampling, redecoration, and KMC graph assembly), and easy swapping or insertion of new modules to support alternative engines or property calculations (Hatton et al., 20 Jun 2025).
• Engine abstraction layers: ASE's calculator interface allows Recipes or nodes to exchange calculators interchangeably, facilitating cross-engine workflows without logic modification (Gjerding et al., 2021, Gelžinytė et al., 2023).

Extensibility is further supported by plug-in architectures (CatLearn fingerprint, regressor, and acquisition function APIs (Hansen et al., 2019)), minimal effort ports to new reaction types (WhereWulff reaction-scheme plugin (Sanspeur et al., 2023)), and semantic orchestration layers (semantic workflow chaining in pyiron/ConceptualDict (Guzman et al., 1 Feb 2026)).

4. Workflow Orchestration, Resource Management, and Automation

Workflow orchestration encompasses dependency resolution, parallel task scheduling, error handling, migration, and runtime adaptation:

• DAG-based orchestration: Recipes, Instructions, or workflow nodes are composed into DAGs, with cached Records or Results enforcing full dependency-resolved execution (ASR Workflow, AiiDA DAG, MiqroForge visual editor, HopDec module graph).
• Automated job management: Integrations with resource managers (SLURM, PBS, LSF, Fireworks, MyQueue, ExPyRe) allow scaling from single-node to high-throughput or HPC systems, with robust error recovery and checkpoint/restart protocols (WhereWulff's ContinueOptimizeFW, ASR's MyQueue, wfl's ExPyRe) (Gjerding et al., 2021, Gelžinytė et al., 2023, Sanspeur et al., 2023).
• Resource-aware scheduling: Advanced frameworks like MiqroForge employ AI-driven scheduling, with dynamic provisioning based on node performance profiles and real-time DAG analysis (Wang et al., 11 Aug 2025).
• Migration and update propagation: Workflow/data migration modules enable propagation of bugfixes and metadata changes to cached or completed Records without recomputation, ensuring long-term maintainability (ASR Migration module).

5. Semantic Enrichment, Knowledge Representation, and AI Integration

Knowledge-based workflows leverage semantic metadata and knowledge graphs to enable machine-understandable representations and agentic automation:

• Ontology-driven metadata annotation: Each workflow node outputs detailed metadata mapped to ontology concepts (ASMO for simulation methods, CMSO for sample/material, PROV-O for agents and activities), enabling automatic provenance capture and semantic validation (Guzman et al., 1 Feb 2026).
• Knowledge graph population: Metadata is serialized to RDF triples and ingested into a knowledge graph, supporting SPARQL queries over property values, method provenance, and material parameters.
• Agentic orchestration: Frameworks such as ChemGraph (Pham et al., 3 Jun 2025) and QUASAR (Yang et al., 30 Jan 2026) coordinate task planning, execution, and aggregation via multi-agent or LLM/AI-driven controllers, with workflows and metadata passed through JSON schemas and message dicts for compositional reasoning and dynamic adaptation.
• AI-ready data outputs: Semantic pipelines provide inputs for retrieval-augmented generation, closed-loop autonomous discovery, and reinforcement learning-driven workflow optimization.

6. Representative Use Cases and Applications

Reusable atomistic workflows have been deployed in a range of contexts:

• High-throughput materials databases: The ASR framework supported the C2DB project, orchestrating ∼60,000 first-principles runs with fully versioned, web-browsable records (Gjerding et al., 2021).
• Quantum/classical hybrid simulations: MiqroForge enables end-to-end QM/MM workflows integrating OpenMM, DFT, and quantum algorithms, with full provenance and containerized node sharing (Wang et al., 11 Aug 2025).
• Defect transport modeling in complex materials: HopDec automates accelerated MD, redecoration analysis, and defect-state graph construction for KMC simulation of chemically complex materials (Hatton et al., 20 Jun 2025).
• FAIR benchmarking and sharing: Sim2Ls on nanoHUB provide modular, parameterized molecular dynamics workflows with validated, unit-aware I/O and global result databases (Hunt et al., 2021).
• Catalyst screening and ML optimization: CatLearn delivers modular ML-accelerated active-learning loops, supporting descriptor selection, regression, and autonomous exploration of catalyst surfaces and nanoparticles (Hansen et al., 2019).
• Semantic simulation and AI-driven property prediction: Semantic workflow frameworks couple pyiron, jobflow, and ontology mapping to enable engine-agnostic computation, cross-method comparison, and agentic workflow construction for elastic, thermodynamic, and defect properties (Guzman et al., 1 Feb 2026).

7. Impact, Limitations, and Future Directions

The emergence of reusable atomistic workflows has led to demonstrable improvements in reproducibility (via provenance), efficiency (via caching and engine abstraction), extensibility (via modular APIs and semantic metadata), and collaborative research (via public registries, FAIR outputs, and workflow migration).

Persistent challenges include:

• Alignment of numerical results across heterogeneous engines, especially for non-pristine or defect-rich structures, due to intrinsic methodological differences (smearing artifacts, symmetry trapping, pseudopotential choices) (Steensen et al., 14 Nov 2025).
• Workflow migration and update propagation in the face of rapid evolution of underlying methods, requiring sophisticated backward-compatibility and migration tooling.
• Semantic interoperability and ontology mapping, which remain bottlenecks for cross-domain data fusion and AI-driven analysis.

Anticipated directions include greater integration of LLM-driven planning and execution, real-time knowledge graph population, richer parameter-space exploration via reinforcement learning, and universal adoption of semantic metadata as a lingua franca for atomistic simulations.

---

Collectively, reusable atomistic workflows represent a maturation of computational materials and chemistry, unifying best practices in reproducibility, provenance, interoperability, and AI readiness (Gjerding et al., 2021, Hunt et al., 2021, Gelžinytė et al., 2023, Pham et al., 3 Jun 2025, Steensen et al., 14 Nov 2025, Yang et al., 30 Jan 2026, Guzman et al., 1 Feb 2026).