FAIR Data Maturity Model
- FAIR Data Maturity Model is a structured framework that converts abstract FAIR principles into assessable, staged criteria for evaluating data and digital objects.
- It employs clear indicators, quantitative metrics, and template-based assessments across metadata, workflows, and AI models to facilitate FAIRification.
- The model supports continuous improvement from raw data to interoperable, machine-actionable assets through both automated checks and community governance.
A FAIR Data Maturity Model is a structured way of describing how well data and related digital objects comply with the FAIR principles—Findable, Accessible, Interoperable, and Reusable—and how that compliance improves over time. In contemporary usage, the scope routinely extends beyond datasets to include AI models, software, notebooks, workflows, documentation, metadata records, and FAIR Digital Objects. Across the literature, the model is not a single standardized instrument but a family of approaches that translate the FAIR principles into observable criteria, metrics, profiles, and process stages suitable for assessment, governance, and continuous improvement (Roy, 2022, Vogt et al., 2024, Santos et al., 2023).
1. Conceptual basis and scope
The core function of a FAIR Data Maturity Model is to convert abstract FAIR guidance into assessable implementation choices. In the FAIR4HEP formulation, this conversion has three components: the four FAIR dimensions are decomposed into specific observable properties; these properties are measured by binary indicators, ordinal levels, or quantitative metrics; and the resulting evidence is organized into levels or stages of maturity (Roy, 2022). This framing is consistent with later work arguing that the FAIR principles are criteria for evaluating data and metadata rather than a standard or a technology, and that FAIRness is not binary but exists on continua (Vogt et al., 2024).
The term “data” is now routinely used in an expanded sense. In high-energy physics, the relevant digital objects include datasets, AI models, software, notebooks, workflows, and documentation (Roy, 2022). In FAIR Digital Object work, the unit of concern is an identified digital object with metadata, type information, and typed relations to other objects (Santos et al., 2023, Blumenroehr et al., 2024). In metadata-template ecosystems, the operative object may be a community-defined metadata template or a FAIR metadata record; in ML infrastructures, it may be a dataset, workflow, execution, or execution asset (Musen et al., 2022, Li et al., 2024). A mature model therefore typically spans both the dataset layer and the semantic, infrastructural, and workflow layers that make datasets operable.
This breadth matters because FAIR implementation is rarely limited to repository deposition. Several works treat FAIRness as a lifecycle property. The ODAM approach for experimental tables in life sciences pushes FAIRification upstream into data preparation and structural metadata design (Jacob et al., 2020). Deriva-ML describes “continuous FAIRness” as making all data FAIR all the time across the investigation lifecycle (Li et al., 2024). The FAIR Funder pilot embeds FAIR requirements at the request-for-proposals stage and follows them through planning, collection, deposition, and automated evaluation (Wittenburg et al., 2019). A plausible implication is that maturity is better understood as a progression of institutionalized practices than as a static score attached to a deposited object.
2. Indicators, metrics, and maturity structuring
Most maturity models operationalize FAIRness by organizing evidence into dimensions, features, and ratings. A catalogue-oriented example represents maturity as a vector
with each dimension decomposed into features,
and assessed through binary presence/absence criteria (Corcho et al., 2023). In that model, the 12 dimensions are Metadata, Openness, Quality, Availability, Statistics, PID, Governance, Community, Sustainability, Technology, Transparency, and Assessment, with 43 associated features (Corcho et al., 2023). This is not a dataset-level FAIR model in the narrow sense, but it shows how maturity can be formalized as a multidimensional profile rather than a single scalar.
Other work uses graded indicators. The assessment of the Digital Shadow Reference Model employs a five-point qualitative scale for each FAIR indicator: --, -, o, +, and ++, ranging from major deficiency to strong alignment with FAIR best practices (Theissen-Lipp, 22 Apr 2025). In that case, 21 questions are mapped to FAIR sub-principles from F1–F4, A1.1–A2, I1–I3, and R1.1–R1.2, covering identifiers, metadata schemas, search indexing, communication protocols, authentication, logical representation, vocabularies, licenses, and provenance (Theissen-Lipp, 22 Apr 2025).
Template-based ecosystems use a different but closely related structure. FAIRware evaluates metadata records against a template-derived gold standard and reports two quantitative dimensions: completeness, defined by the presence or absence of required values in the metadata fields defined in the metadata schema, and adherence, defined by whether the filled values conform to the field’s data type and ontology constraints (Musen et al., 2022). In practical terms, this means that maturity can be assessed record by record as the proportion of required fields filled and the proportion of filled fields that are valid relative to community standards.
Several papers also describe maturity-like levels explicitly or implicitly. FAIR4HEP reconstructs a dataset curve from “Level 0 – Raw, experiment-local data” through “Level 2 – FAIR and AI-ready for broad ML community,” and a model curve from “Level 0 – Prototype code only” to “Level 2 – Interoperable, performance-characterized model” (Roy, 2022). FAIR 2.0 proposes a progression from machine-readable to machine-interpretable to machine-actionable, and from syntactic handling to advanced semantic interoperability (Vogt et al., 2024). The parametric review of 13 FAIR implementation frameworks suggests another layering: awareness of FAIR, planned FAIRification, implemented FAIR processes, and optimized FAIR ecosystems that combine technical tooling with governance, policy, and training (Singh et al., 2024).
3. Semantic interoperability as an advanced maturity dimension
A major expansion of the FAIR maturity concept is the move from syntactic compliance toward semantic interoperability. FAIR 2.0 defines semantic interoperability as the condition in which transferred information has, in its communicated form, all of the meaning required for the receiving system to interpret it correctly, and decomposes it into terminological and propositional interoperability (Vogt et al., 2024). This has direct consequences for maturity modeling.
Terminological interoperability concerns the alignment of terms. FAIR 2.0 distinguishes ontological interoperability, where terms have the same meaning and referent, from referential interoperability, where terms share the same referent or extension but may differ in meaning (Vogt et al., 2024). The corresponding maturity indicators are the use of FAIR vocabularies, the existence of ontological and referential entity mappings, the use of standards such as SSSOM, and multilingual labels and synonyms. Propositional interoperability concerns statements and schemas. It requires either the same schema for the same type of statement or formal schema crosswalks across different schemata, together with explicit logical frameworks and statement typing (Vogt et al., 2024).
This semantic turn changes the maturity target. Earlier FAIR assessment tools such as FAIR-Checker, F-UJI, and FAIR Evaluator have mostly been applied to basic provenance and licensing metadata, not to domain-specific data semantics (Vogt et al., 2024). FAIR 2.0 argues that maturity must also cover schema identifiers, schema crosswalks, logical framework declarations, statement distinctions such as lexical, assertional, contingent, prototypical, and universal, and certainty metadata (Vogt et al., 2024). A plausible implication is that mature FAIRness requires explicit machine-understandable mappings and crosswalks, not merely well-formed metadata.
The same point appears in infrastructure studies. The maturity model for catalogues of semantic artefacts treats ontologies, vocabularies, mappings, standards, and schemas as foundational infrastructure for FAIR semantics and evaluates whether they are persistent, governed, queryable, aligned, and sustainable (Corcho et al., 2023). YAMZ extends this to community-driven vocabulary governance by assigning ARKs to terms, storing RDF subject–predicate–object relationships, and moving terms through the states Vernacular, Canonical, and Deprecated based on consensus scoring and stability (Rauch et al., 2021). In maturity terms, these works shift attention from datasets alone to the semantic substrate that makes data interoperable and reusable.
4. Infrastructure, templates, and FAIR Digital Objects
A recurring theme in the literature is that maturity depends on reusable infrastructure components rather than isolated local practice. One such component is the machine-actionable metadata template. In the CEDAR and FAIRware ecosystem, communities translate reporting guidelines and metadata standards into JSON Schema–based templates and JSON-LD metadata instances, with ontology-linked fields and explicit constraints (Musen et al., 2022). These templates act as a community reference for what constitutes rich and domain-relevant metadata, and FAIRware evaluates archived datasets against them by checking missing required values, datatype conformance, and ontology-term validity (Musen et al., 2022). This makes template availability itself a maturity criterion.
Another component is workflow-aware planning and policy infrastructure. The FAIR Funder pilot envisions a seven-stage data-management workflow in which Metadata for Machines workshops define community-specific metadata schemas and FAIR metrics; templates and metrics are stored in CEDAR and FAIRsharing; funders compose call requirements and embed them in Data Stewardship Wizard; researchers and institutional data stewards create machine-actionable data stewardship plans; data are collected with FAIR-supportive tools such as CEDAR and Castor EDC; and repositories trigger third-party FAIR metrics evaluation and certification (Wittenburg et al., 2019). This model makes maturity a property of the funding and stewardship ecosystem, not only of a dataset.
FAIR Digital Object work supplies a further infrastructural layer. The ontology-driven conceptual model for FAIR Digital Objects defines an FDO as an identified object with a Globally Unique, Persistent, Resolvable Identifier, at least one FAIR Metadata Record, and an explicit distinction between informational objects and their materializations (Santos et al., 2023). The more formalized 2024 model specifies that an FDO has a unique PID, exactly one information record instantiating exactly one Kernel Information Profile, a mandatory kernel attribute set including KIP reference, License, Checksum, Digital Resource Location, Creation Date, and Digital Resource Type, and typed links to other entities that form a PID-graph (Blumenroehr et al., 2024). In maturity terms, this yields concrete assertions that can be automatically tested.
The infrastructural literature also shows that maturity has a strong organizational aspect. The parametric review of implementation frameworks finds that most frameworks are technology-first, with less consideration for the social aspects of FAIR, and argues for a people-first, human-centered design approach (Singh et al., 2024). This suggests that infrastructure maturity is not exhausted by APIs, PID services, or metadata editors; it also includes governance, training, role clarity, policy integration, and feedback mechanisms.
5. Domain implementations
The domain literature shows how FAIR maturity models are instantiated in practice and how their criteria are adapted to local epistemic and technical constraints.
In high-energy physics, the FAIR4HEP collaboration interprets FAIR for both experimental datasets and AI models. For datasets, findability includes DOIs and rich metadata; accessibility includes HTTP/HTTPS retrieval and clear licensing; interoperability includes ML-friendly formats and feature semantics; and reusability includes provenance, educational notebooks, and example analyses (Roy, 2022). For AI models, interoperability is assessed by converting a PyTorch Interaction Network to ONNX and TensorRT, then comparing predictive behavior and performance across frameworks using accuracy, ROC-AUC, inference time per batch, throughput, and GPU utilization (Roy, 2022). The same paper treats cookiecutter4fair and pedagogical notebooks as mechanisms that move projects up a FAIR maturity curve.
In scientific AI more broadly, FAIR principles for AI models are made explicitly measurable as pass/fail conditions. A model is findable when a DOI resolves to rich model metadata and executable artifacts; accessible when it is downloadable or invocable via standard network protocols and its metadata persist; interoperable when it uses standard metadata formats and containerization across hardware architectures; and reusable when humans or machines can reproduce its advertised capabilities using published validation data and uncertainty or performance metrics (Ravi et al., 2022). This extends data maturity models into data–model ecosystems and makes execution environment, containerization, and hardware portability part of advanced FAIRness.
In life-science experimental tables, the ODAM approach shows a stepwise FAIRification path from ordinary spreadsheets to structurally annotated, ontology-linked datasets that can be assessed with the RDA FAIR Data Maturity Model, OZONOME, and FAIR SHARC (Jacob et al., 2020). The key operational devices are one-entity-per-file organization, identifier columns such as PlantID and SampleID, attributes categorized as identifier, factor, quantitative, or qualitative, and structural metadata exported as a Frictionless datapackage (Jacob et al., 2020). This is a highly concrete example of a maturity model where structural organization and semantic enrichment are successive layers.
In computational materials mechanics, a workflow-centric schema treats each workflow run as a unique data object combining user-specific elements, system elements, job-specific elements, and property elements (Shoghi et al., 2024). The schema records geometry, boundary conditions, constitutive model parameters, texture, stress–strain outputs, and units, and is explicitly designed to generate FAIR data objects for microstructure-sensitive mechanical data (Shoghi et al., 2024). In Deriva-ML, the same lifecycle logic is applied to ML investigations: datasets, workflows, executions, execution assets, and execution metadata are cataloged in an entity–relationship model, large datasets are packaged as BDBags with Minids, and FAIRness is treated as a continuous property of the investigation rather than a terminal deposition event (Li et al., 2024).
6. Limits, critiques, and future directions
The main critique of FAIR Data Maturity Models is that FAIR compliance is not identical to practical usability. The Dataset Friction Framework argues that FAIR is, by design, a stewardship framework describing dataset properties and metadata, whereas it does not describe the effort a specific user must expend to move from awareness of a dataset to productive use (Pidduck et al., 22 Jun 2026). DFF therefore measures user-facing friction across six dimensions—Discoverability and Understanding, Access and Delivery, Licence and Legal, Data Structure and Format, Tooling and Support, and Overall Complexity—and distinguishes engineered friction from accidental friction (Pidduck et al., 22 Jun 2026). Its central claim is that FAIR compliance and DFF friction are non-redundant and jointly informative, so maturity models that equate FAIR with usability risk systematic mismeasurement.
A second critique concerns overemphasis on technology. The parametric review of implementation frameworks concludes that most frameworks are very technical in nature and seem to be adopting the technology-first approach, while missing critical aspects of explaining what, why, and how for the four foundational principles of FAIR and giving less consideration to social aspects (Singh et al., 2024). This suggests that mature FAIR implementation must include competencies, governance, incentives, institutional roles, and policy integration, not only metadata schemas and tool deployment.
A third limitation is the difficulty of full automation. FAIR 2.0 notes that semantic interoperability depends on mappings, schema crosswalks, logical frameworks, and statement typing that current automated FAIR assessments generally do not evaluate (Vogt et al., 2024). The template literature similarly shows that computational evaluation is only reliable when community beliefs about richness and domain relevance have first been codified in machine-actionable templates (Musen et al., 2022). A plausible implication is that mature assessment ecosystems require both automation and community governance over the standards that automation consumes.
Future directions in the literature are correspondingly multi-layered. FAIR 2.0 points toward combined FAIR plus semantic-interoperability levels, enriched by terminology, schema, and operations services (Vogt et al., 2024). FAIR Digital Object work points toward globally aligned data spaces built on domain-independent abstraction, encapsulation, and typed entity relationships (Blumenroehr et al., 2024). The FAIR Funder pilot points toward certification schemes for datasets, repositories, practitioners, and organizations (Wittenburg et al., 2019). Across these strands, the FAIR Data Maturity Model is moving from a repository-oriented checklist toward a broader architecture for semantically explicit, machine-actionable, workflow-aware, and institutionally governed data ecosystems.
