Knowledge Infrastructure (KI)
- Knowledge Infrastructure (KI) is a socio-technical system that integrates people, artifacts, and institutions to generate, curate, and sustainably share knowledge.
- It operationalizes knowledge through layered technical, organizational, and epistemic components, enabling automated extraction, curation, and execution.
- KI enhances interdisciplinary collaboration by establishing standards, governance practices, and human–machine co-production for reliable, reusable outputs.
Knowledge infrastructure denotes a socio-technical arrangement through which knowledge is generated, shared, maintained, and rendered usable across time. In the contemporary literature, the canonical definition describes knowledge infrastructures as “robust networks of people, artifacts, and institutions that routinely generate, share, and maintain specific knowledge about the human and natural worlds,” with robustness and endurance treated as constitutive rather than optional properties (Gregory et al., 26 Feb 2025). Current work applies this concept to scholarly communication, science communication, engineering, Earth-system simulation, and public epistemic systems, while emphasizing that a knowledge infrastructure is not merely a repository or platform but an organized combination of standards, tools, governance, workflows, and communities (Jaradeh et al., 2019, Wittenborg, 6 Mar 2026, Li et al., 18 May 2026, Gardiner, 12 Jun 2026).
1. Definition, scope, and conceptual lineage
The contemporary concept of knowledge infrastructure is anchored in infrastructure studies, especially Edwards’ formulation of infrastructures as robust networks of people, artifacts, and institutions. In this framing, infrastructures are socio-technical, historically embedded, and relational. They include technical components such as repositories, standards, tools, and platforms; human actors such as researchers, curators, citizens, and managers; and organizational arrangements such as host institutions, governance bodies, and funding regimes (Gregory et al., 26 Feb 2025).
A central implication of this definition is that sustainability is internal to the concept itself. Robustness means being “capable of withstanding the test of time,” so persistence is part of what qualifies an arrangement as a knowledge infrastructure at all (Gregory et al., 26 Feb 2025). This aligns with the distinction between “project time” and “infrastructure time”: a project may build a useful artifact, but an infrastructure must endure beyond the lifetime of the originating project and support continuing design, maintenance, and reuse (Wittenborg, 6 Mar 2026).
Recent literature also broadens the referent of knowledge infrastructures beyond disciplinary science. Earlier work often treated them as infrastructures serving scientific communities; newer accounts emphasize that they increasingly serve “researchers, citizens, policymakers, and others,” especially under conditions of open science, citizen science, digital justice, and platformization (Gregory et al., 26 Feb 2025). This expansion is visible in work on science communication, where videos and podcasts become knowledge objects, and in Earth-science simulation, where infrastructures are designed to make expert modelling practice accessible to non-specialists (Wittenborg et al., 12 Nov 2025, Li et al., 18 May 2026).
The acronym “KI” is polysemous. In parts of machine-learning and NLP research it denotes unrelated notions such as “knowledge inheritance,” “knowledge integration,” or “knowledge-intensive” tasks; those usages concern model training or evaluation rather than Edwards-style infrastructures and should be distinguished from the infrastructure concept (Qin et al., 2021, Hou et al., 2022, Piktus et al., 2021).
2. Constituent elements and infrastructural forms
Across domains, knowledge infrastructures are described as layered combinations of technical substrates, organizational forms, and epistemic practices. Technical layers include repositories, knowledge graphs, wiki systems, APIs, standards, exchange formats, semantic schemas, digital libraries, and workflow tools. Organizational layers include roles, routines, communities, institutions, and governance mechanisms. Epistemic layers include curation criteria, provenance rules, validation procedures, and maintenance practices (Jaradeh et al., 2019, Wittenborg, 6 Mar 2026).
Several papers make this layered structure explicit. The Open Research Knowledge Graph represents scholarly knowledge in a layered architecture comprising a persistence layer with a labeled property graph, an RDF triple store, and relational components; a domain model centered on ResearchContribution; and service layers for contribution, curation, and exploration exposed through a REST API (Jaradeh et al., 2019). Aerospace.Wikibase similarly combines an OWL domain model, a Wikibase knowledge graph, and a Python-based ingestion and maintenance layer, with a SPARQL endpoint and stable IRIs for items and properties (Wittenborg et al., 5 Mar 2026).
The literature also distinguishes different residences and governance forms of knowledge-bearing stock. A knowledge theory of capital identifies embodied knowledge, disembodied knowledge, institutionalized knowledge, commons knowledge, and public epistemic infrastructure as analytically distinct forms. In that account, public epistemic infrastructure includes universities, public research institutions, statistical agencies, standards, legal doctrines, and peer-review systems, while commons knowledge includes open-source projects and shared data resources (Gardiner, 12 Jun 2026). This suggests that “knowledge infrastructure” is not reducible to software or data alone; it also includes the human and institutional capacities that make those artifacts productive.
Older domain-specific work anticipated this broadened view even where the term “knowledge infrastructure” was not explicit. In medical science, the relevant infrastructure is described through knowledge management systems, knowledge bases, portals, discussion forums, telemedical consultation, security mechanisms, and organizational learning practices (Bohlouli et al., 2020). In engineering design, a standards-based infrastructure for knowledge representation and exchange is built from UML, OCL, XML, XMI, RDF, and RDFS, with ontologies and web architecture providing semantic and transport layers (Buzon et al., 2018).
3. Core operations: representation, curation, execution, and reuse
Knowledge infrastructures do not merely store artifacts; they operationalize knowledge. A recurrent motif is the shift from document- or file-centric organization to machine-actionable, executable, or agent-actionable forms. In ORKG, this takes the form of structured ResearchContribution graphs linking problems, methods, and results, enabling comparison, exploration, export, and reuse beyond bibliographic metadata (Jaradeh et al., 2019). In science communication, the objective is FAIR media representation for videos and podcasts through structured entities, collaborative annotations, and integration with full-text artifacts such as transcripts (Wittenborg et al., 12 Nov 2025).
A second recurrent operation is human–machine co-production. ORKG combines crowdsourced knowledge acquisition by authors, reviewers, librarians, and curators with automated extraction from scholarly texts, including named entity recognition and entity linking (Jaradeh et al., 2019). The doctoral work on fostering knowledge infrastructures generalizes this pattern into human-in-the-loop workflows for systematic literature reviews, knowledge extraction, and downstream knowledge-graph construction, notably through SWARM-SLR and ExtracTable (Wittenborg, 6 Mar 2026). This literature treats automation not as a replacement for curation but as a way to reduce friction in otherwise manual, high-volume workflows.
A third operation is executable use. “Networks of knowledge engines” recast expertise as API-accessible software services that automatically generate problem-specific outputs and can be composed into larger solution-creation systems (Bergmair et al., 2018). KISS pushes this logic further for scientific simulation by externalizing operational expertise into validated modelling operators, staged domain protocols, and diagnostic recovery mechanisms so that agents can run process-based models correctly (Li et al., 18 May 2026). The resulting infrastructure is not just searchable knowledge but callable, staged, and recoverable knowledge.
These developments indicate a broad shift in the unit of infrastructure content. The unit may be a research contribution, a media item, a modelling operator, an ontology-backed concept, or an executable knowledge engine. The common denominator is that the infrastructure captures enough structure for computational systems and human communities to reuse knowledge without repeatedly reconstructing it from scratch.
4. Domain implementations
The concept is instantiated differently across domains, but several exemplars recur in the recent literature.
| Domain | Infrastructure | Distinctive focus |
|---|---|---|
| Scholarly communication | ORKG | Machine-actionable research contributions |
| Science communication | SciCom Wiki | FAIR representation of videos and podcasts |
| Aerospace engineering | Aerospace.Wikibase | Shared taxonomy and identifiers for processes, software, and data |
| Earth-science simulation | KISS / KDT | Agent-actionable operators, protocols, and recovery mechanisms |
| Distributed expertise | Networks of Knowledge Engines | Executable knowledge services and runtime composition |
In scholarly communication, ORKG is explicitly presented as a “next generation infrastructure for semantic scholarly knowledge.” It moves beyond PDFs plus bibliographic metadata by representing problems, methods, data, and results in a structured knowledge graph, supporting comparison tables, similarity search, and machine-actionable curation (Jaradeh et al., 2019).
In science communication, SciCom Wiki is designed as an open, Wikibase-centered digital library for scientific videos and podcasts. Its purpose is to provide a central, collaborative node for discovery, annotation, and FAIR metadata capture in a domain that the literature describes as fragmented and underdeveloped (Wittenborg et al., 12 Nov 2025). The related doctoral work places this system in a broader “SciCom KI,” a network of people, artifacts, and institutions around science communication media (Wittenborg, 6 Mar 2026).
In aerospace engineering, Aerospace.Wikibase is explicitly framed as a knowledge infrastructure for a domain whose knowledge is often project-bound, closed, and discontinuous. Built on Wikibase and initially populated from a systematic literature review, it provides over 700 terms related to processes, software, and data, and links project-independent concepts to persistent, open infrastructure without requiring disclosure of project-specific information (Wittenborg et al., 5 Mar 2026).
In Earth-science simulation, KISS treats KI as an agent-actionable scaffold around process-based models. In a 3,000-trial coupled-hydrology benchmark, agents equipped with KI produced physically plausible, verifiable end-to-end simulations in up to 84% of trials, whereas agents without KI plateaued below 40%. Its Knowledge Dissection Toolkit then generated KI for 117 additional process-based models across 14 Earth-science domains, supporting a total corpus of 119 KIs (Li et al., 18 May 2026).
A more general infrastructural model appears in “Networks of Knowledge Engines,” where expert knowledge is encapsulated in automated, API-accessible engines and composed into executable networks at runtime. This extends the infrastructure concept from knowledge storage toward scalable, on-demand deployment of expert competence (Bergmair et al., 2018).
5. Governance, sustainability, and evaluation
A major theme in the recent literature is that sustaining a knowledge infrastructure is an ongoing governance problem rather than a one-time technical achievement. Gregory and colleagues argue that sustainability is never “solved” because technologies, standards, funding models, user needs, and contexts continually evolve. They organize this problem through five recurring questions: when to talk about sustainability, how to communicate value, how to bring communities together, what the right size of an infrastructure is, and how to make sustainability decisions in an ongoing reflective way (Gregory et al., 26 Feb 2025).
This literature treats governance, evaluation, and legitimacy as infrastructural, not peripheral. Persistent identifiers, metrics, and logic models are presented as part of the machinery by which infrastructures demonstrate value and secure continued support (Gregory et al., 26 Feb 2025). Governance principles such as stakeholder participation, transparency, interoperability, and adaptability are highlighted in POSI and GORC, which are discussed as sustainability levers rather than mere normative add-ons (Gregory et al., 26 Feb 2025).
Empirical evaluations in specific systems reinforce the socio-technical character of KI. ORKG’s first user study, conducted with 12 conference authors, reported that 75% found the front end “fairly intuitive and easy to use,” the average completion time for creating a structured contribution was 17 minutes, and participants created on average about 53 triples per contribution (Jaradeh et al., 2019). SciCom Wiki was designed from requirements elicited from 53 stakeholders, refined in 11 interviews, and evaluated with 14 additional participants; the system was judged suitable for improving discovery and curation, while legal, onboarding, and scalability work remained necessary for broader adoption (Wittenborg et al., 12 Nov 2025). In KISS, evaluation is operational rather than survey-based: KI-equipped agents substantially outperform agents without KI on end-to-end simulation tasks, which functions as a direct test of whether externalized operational knowledge improves reliable use (Li et al., 18 May 2026).
The sustainability literature also rejects simple growth narratives. “Right-sizing,” degrowth, and planned sunsetting are treated as legitimate infrastructural strategies when they preserve relevance and capacity better than indiscriminate expansion (Gregory et al., 26 Feb 2025). This is especially pertinent where infrastructures are vulnerable to project-based discontinuity, as in aerospace and science communication (Wittenborg et al., 5 Mar 2026, Wittenborg, 6 Mar 2026).
6. Challenges, controversies, and future directions
Fragmentation is the dominant problem across the surveyed domains. In science communication and aerospace engineering, knowledge remains dispersed across formats, repositories, and organizational boundaries, reducing findability, accessibility, interoperability, and reusability (Wittenborg, 6 Mar 2026). In aerospace this fragmentation coexists with strong technical standards such as CPACS and CMDOWS, but adoption is uneven and constrained by intellectual property, regulation, and institutional secrecy (Wittenborg et al., 5 Mar 2026, Wittenborg, 6 Mar 2026). In science communication, the challenge is less legal closure than the absence of a scalable, shared structure for non-textual media and their annotations (Wittenborg et al., 12 Nov 2025).
A second persistent tension concerns openness versus control. Public epistemic infrastructures and commons are repeatedly identified as foundational, yet they are vulnerable to underfunding, enclosure, or platform capture (Gardiner, 12 Jun 2026). The knowledge theory of capital sharpens this by distinguishing first conversion, cognitive enclosure, feedback capture, dark capital, and expected knowledge loss. In that framework, software, data, models, organizations, commons, and public epistemic infrastructure all function as knowledge-bearing stock, and modern wealth depends not only on accumulating such stock but on how it is governed (Gardiner, 12 Jun 2026). A plausible implication is that KI research increasingly intersects with political economy: the problem is not just how to build infrastructures, but how to keep their generative capacities from being narrowed by enclosure, dependency, or maintenance deficits.
A third challenge is granularity. ORKG deliberately adopts a simple problem–method–result model to lower barriers to contribution, while acknowledging that richer semantics are desirable (Jaradeh et al., 2019). KISS shows that operational expertise can be decomposed into recurring categories of decisions and failure remedies across different scientific domains, but its authors also note that validation depth, domain scope, and long-term maintenance remain open issues (Li et al., 18 May 2026). Similar trade-offs appear in SciCom Wiki, where support for rich annotation, accessibility, trust, and fact-checking must be balanced against usability and scale (Wittenborg et al., 12 Nov 2025).
Future work in this area consistently moves in two directions. One is deeper machine actionability: more automated extraction, better interoperable schemas, stronger links between knowledge graphs, digital libraries, and executable services, and infrastructures that can serve both humans and agents (Jaradeh et al., 2019, Li et al., 18 May 2026). The other is deeper socio-institutional embedding: sustained governance, community uptake, legal accommodation, and infrastructures that endure beyond project cycles and remain answerable to the publics they serve (Gregory et al., 26 Feb 2025, Wittenborg, 6 Mar 2026). Across these strands, the central claim is stable: knowledge infrastructure is the organized means by which knowledge becomes durable, shareable, governable, and reusable at scale.