Semantic Framework Enabling Machine Learning in Manufacturing
Arild Waaler1, Martin G. Skjæ veland¹ and Dimitris Kyritsis¹
1 Department of Informatics, University of Oslo, Norway
E-mail: arild@uio.no
Status
Engineering practice today is fundamentally dependent on technical information, information that is document-based, fragmented across tools and domains, and tied to siloed organizational structures. While individual disciplines to varying degrees employ formal models like ontologies, there is no shared foundation for structuring system-level knowledge across the engineering lifecycle. As a result, the so- called digital thread — the traceable connection from requirements through design, implementation, and operation — is frequently broken or opaque.
At the same time, artificial intelligence, particularly in the form of machine learning and generative models, is rapidly being integrated into engineering workflows. However, these approaches struggle to deliver trustworthy results in the absence of clearly defined objects and relations. Where structure is implicit or missing, AI models become unreliable, and engineers are left without the means to validate or interpret the outputs of AI models. This creates a gap between the promise of AI and the reality of high-stakes engineering practice, where precision, safety, trust, and traceability are imperative.
To address this, there is an urgent need for structured representations of engineering knowledge that are both verifiable and understandable. This requires enabling expert users — across engineering disciplines — to validate, reuse, and refine the information and data models that underpin AI and automation. Crucially, such validation should be grounded in established systems engineering principles: abstraction for information hiding, hierarchical decomposition for modularity, topology for managing flows, interfaces for encapsulation, and classification for reuse.
Formal verification demands even more: a logical foundation that supports tractable reasoning. This includes the ability to define and check class axioms, detect inconsistencies, and infer consequences within decidable subsets of logic. Without such foundations, technical information models cannot be reliably queried, reused, or integrated at scale.
The lack of structure in technical information also undermines efforts to build scalable industrial knowledge graphs. While knowledge graph and ontology standards such as RDF [8] and OWL [9] are widely used, they often lack connection to engineering practice and do not capture the structural logic of systems. A principled foundation is needed, one that connects engineering semantics with semantics-based representations and can serve both as a modeling framework and as machine-readable input to AI pipelines [1].
Current and future challenges
To enable digital transformation in engineering, we need languages that allow engineers to describe systems in a way that is both human-readable and machine-actionable. Such languages must support abstraction, modularity, encapsulation of interfaces, and reuse — all core principles of systems engineering. They must also allow engineers to express partial, evolving structures, reflecting the reality that system models are rarely complete at any one time.
At the same time, to support automation, AI, and formal reasoning, these languages must have a well-defined logical foundation. Ontology languages like OWL offer precise semantics based on Description Logic [10], supporting classification, consistency checking, and inference. However, they are often ill-suited to capture the structural and contextual richness of engineering systems: they lack native support for system-level constructs such as breakdown hierarchies, connectivity relations, modalities like intended versus actual configurations, and lifecycle-specific views.
This reveals a fundamental conceptual gap. Engineering requires modelling languages that express intensional structure — definitions of system elements in terms of their roles, constraints, and relationships — while ontology-based approaches typically focus on extensional classification and static taxonomies. Current tools and languages seldom support the coherent expression of intensions in a way that can be incrementally developed, reused, and verified.
This gap also manifests in knowledge graph construction: current ontology languages provide semantic rigor, but not the structural expressiveness engineers need to model real systems. Scalable industrial knowledge graphs thus remain difficult to construct and maintain. Moreover, many advanced ML algorithms rely on structured, semantically rich graph inputs — a need yet to be fully met in engineering domains [1].
Bridging these gaps requires rethinking how we represent engineering knowledge: starting from the needs of engineering practice, while grounding models in logic-based semantics and enabling automation. Several key challenges emerge:
-
How can we define structural specifications that are both readable and logically precise?
-
How can we preserve traceability across partial models and evolving designs?
-
How can we align domain-specific engineering practices with shared ontological foundations?
-
How can we combine human-driven design with machine-generated model structures? To meet these challenges, a new framework must treat intensions as first-class citizens, enabling models that are contextual, compositional, semantically transparent, and AI-ready.
Advances in science and technology to meet challenges
Addressing the gap between engineering modelling needs and formal ontology capabilities requires a new kind of framework, one that combines the expressive power of systems engineering with the precision of formal logic. The Information Modelling Framework (IMF) [7, 2] is designed to meet this need. It provides a simple yet comprehensive and extensible core language that allows engineers to express intensional definitions of system elements in a way that is formally interpretable, incrementally buildable, and amenable to verification.
The foundation of IMF is a clear separation between intension and extension. An IMF element expresses a structural specification, a parameterized definition of what a system element is intended to be, rather than what currently exists. Each such specification is annotated with an aspect: a structured context capturing the information domain (e.g., function, implementation), the modality (e.g., intended, actual), and the lifecycle perspective (e.g., product, project), see Figure 1. This framing enables coherent modeling of alternative designs, requirements vs. implementations, and different stakeholder views — all within a unified formalism.
| Figure 1. | Information in the Information Modelling Framework (IMF) is represented in distinct and | |
|---|---|---|
| interrelated aspects. The figure illustrates four complementary specifications of different aspects that together provide a comprehensive description of a pump: | intended functional requirements (1, yellow), intended | |
| implementation (product) specification | (2, cyan), | intended spatial requirements (3, magenta), and actual |
| installed descriptions | (4, blue). |
Formally, IMF uses constructs from typed lambda calculus to define structural specifications. Relations such as hasPart, connectedTo, and hasTerminal are interpreted as function applications, with semantics grounded in intensional logic. The resulting models can be translated into extensional axiom sets (e.g., in Description Logic), enabling formal reasoning tools to verify properties, detect inconsistencies, and support model completion.
Because IMF models are grounded in formal semantics and can be serialized as knowledge graphs using RDF, they serve as a natural foundation for industrial knowledge graphs. These graphs are semantically rich, structured according to systems engineering logic, and readable by machines. Moreover, the RDF representation of IMF models can be directly consumed by advanced ML pipelines, providing AI models with well-formed, semantically validated engineering structures.
Figure 2. In the EU-funded project Tec4MaaSEs [6], the Information Modelling Framework (IMF) is used to
capture and represent both the requirements for complex engineering artefacts and the specifications of the equipment intended to fulfil them. Technical information from various perspectives and in different formats (depicted on the left) is consolidated within a structured IMF information model (top right). This model serves as the foundation for generating an RDF knowledge graph and an OWL ontology (bottom right), enabling automated verification, reasoning, and analysis.
IMF is defined in an openly accessible specification [7] and supported by a Recommended Practice for Asset Information Modelling [2]. Furthermore, the IMF program publishes its semantic technology resources, including its OWL ontology and SHACL shape patterns [3]. IMF is not a fixed standard, but a platform for structured, formal, and open-ended model development. It is already being explored in the EU-funded projects RE4DY [4], SM4RTENANCE [5], and Tec4MaaSEs [6] — see Figure 2 for an example.
Concluding remarks
The Information Modelling Framework (IMF) offers a principled foundation for structuring engineering knowledge in ways that support human understanding, AI-driven automation, and formal verification. By bridging the gap between system modelling practices and logic-based ontologies, IMF enables a new generation of engineering tools that are modular, semantically precise, and adaptable across contexts.
Crucially, IMF provides a foundation for scalable industrial knowledge graphs. Its RDF serialisation format allows both semantic integration and direct input to machine learning workflows — a key enabler for engineering AI.
More than a static language, IMF is a research and innovation program. It invites collaboration from engineers, logicians, data scientists, and tool developers to extend its capabilities and apply it to real-world challenges. By combining systems principles with formal semantics, IMF supports the development of trustworthy, explainable, and scalable digital engineering infrastructure.
Acknowledgements
This research has been partially supported by the European Commission in the Horizon Europe projects RE4DY (grant 101058384), SM4RTENANCE (grant 101123490) and Tec4MaaSEs (grant 101138517).
References
[1] Waaler A and Kiritsis D 2025 Information Modelling Framework for Digital Engineering Digital Engineering 4 100042 doi:10.1016/j.dte.2025.100042
[2] DNV 2024 Asset Information Modelling Framework: Structuring Digital Assets DNV-RP-0670 Online: https://www.dnv.com/digital-trust/recommended-practices/asset-information-modelling-dnv-rp-0670/ (accessed June 2025)
[3] IMF Programme IMF Semantic Technology Resources. 2025. Online: http://ns.imfid.org (accessed June 2025)
[4] RE4DY Project 2025. Online: https://re4dy.eu (accessed June 2025)
[5] SM4RTENANCE Project 2025. Online: https://sm4rtenance.eu (accessed June 2025)
[6] Tec4MaaSEs Project 2025. Online: https://tec4maases.eu (accessed June 2025)
[7] IMF Programme 2025 IMF Overview and Manual Online: https://www.imfid.org/ (accessed June 2025)
[8] W3C. RDF 1.1 Primer. 2014. https://www.w3.org/TR/rdf11-primer/ (accessed June 2025)
[9] W3C. OWL 2 Web Ontology Language Document Overview (Second Edition). 2012. https://www.w3.org/TR/rdf11-primer/ (accessed June 2025)
[10] Franz Baader, Ian Horrocks, Carsten Lutz, Uli Sattler, An Introduction to Description Logic. Cambridge University Press.