What Is Data Curation?

Data curation is a discipline concerned with the active management of data throughout its lifecycle to ensure it remains accurate, accessible, and reusable over time. It encompasses the processes of organizing, describing, cleaning, enriching, and preserving data so that researchers, engineers, and organizations can reliably build on it. The term originated in library and archival science but has expanded into scientific research infrastructure and, more recently, machine learning, where the quality of training datasets has proven to be as consequential as model architecture.

Data curation is distinct from data collection and data analysis: collection creates the raw resource, analysis extracts insight from it, and curation maintains the integrity of the resource itself between those two activities. As datasets grow in scale and are reused across teams and over years, the curatorial layer becomes the difference between data that accumulates trust and data that accumulates technical debt.

Data Quality and Cleaning

The central technical task in data curation is assessing and improving data quality along several dimensions: accuracy (values reflect the real-world entities they represent), completeness (required fields are populated), consistency (values conform to the same format and semantics across records), and timeliness (records are current). Data cleaning addresses identified deficiencies through operations such as deduplication, outlier detection, imputation of missing values, and normalization of inconsistent representations. The FAIR Data Principles, developed through European research infrastructure initiatives and documented by GO FAIR, establish the framework under which curated data should be Findable, Accessible, Interoperable, and Reusable, and many curation workflows are organized around achieving these properties explicitly.

Metadata and Documentation

Metadata is the structured description that allows data to be discovered, understood, and reused without requiring contact with the original creators. A curated dataset carries metadata at multiple levels: descriptive metadata (what the data is about, who created it, when, and why), structural metadata (how fields relate to one another and what units or controlled vocabularies apply), and provenance metadata (what transformations, filters, and quality checks the data has undergone). Metadata schemas vary by domain: the Dublin Core standard covers general digital resources, while domain-specific schemas like ISO 19115 for geospatial data and DICOM for medical imaging encode field-level semantics that generic schemas cannot represent. The Digital Curation Centre's guidance on data management and curation practice outlines how metadata and documentation practices apply across scientific and engineering data contexts.

Long-Term Preservation

Data preservation extends curation across time horizons measured in years or decades. It addresses format obsolescence (proprietary file formats may become unreadable as software changes), storage medium degradation (bit rot and media failure), and organizational continuity (data repositories must outlast the projects that created them). Trusted digital repositories apply practices drawn from the Open Archival Information System (OAIS) reference model, an ISO standard (ISO 14721) that defines the ingest, archival storage, and dissemination functions of a long-lived repository. Preservation decisions include choosing open, well-documented file formats, maintaining multiple geographically distributed copies, and performing periodic integrity checks against recorded checksums. The OSTI.gov scientific and technical information repository managed by the U.S. Department of Energy is an example of a long-term preservation infrastructure for research data and publications produced by federally funded research.

Applications

Data curation has applications in a wide range of disciplines, including:

• Scientific research data management, where curated datasets enable reproducibility and citation
• Machine learning dataset preparation, where curation of training data directly affects model performance and bias
• Clinical and genomic databases, where patient-level data requires curation for regulatory compliance and research reuse
• Engineering simulation archives, where preserving validated datasets supports future design work
• Open government data portals, where curation enables public access to administrative and statistical records