I40 Core Information Model Overview
- I40 Core Information Model is a canonical conceptual framework that underpins manufacturing ontologies by defining structured classes, properties, and constraints.
- The model enables the translation of natural-language manufacturing queries into SPARQL over knowledge graphs by providing context-rich schema annotations via its extension, CIMM.
- Its design separates logical and physical domains, addressing semantic misalignments and improving accuracy in LLM-mediated query synthesis as evidenced by detailed experimental evaluations.
The I40 Core Information Model, also written I4.0 Core Information Model or I40 CIM, is presented as the canonical Industry 4.0 domain model that underpins community manufacturing ontologies and provides a shared, structured conceptual schema for manufacturing knowledge graphs (Monka et al., 30 Jul 2025). In the reported study, it is not queried directly. Instead, it appears through the Core Information Model (CIMM), which is described as “a rich extension of the well-known I40 Core Information Model,” “based on the international IEC manufacturing standard,” and used as the concrete ontology for experiments that translate natural-language manufacturing questions into SPARQL. Within that setting, I40 CIM functions as the conceptual foundation for modeling core manufacturing concepts across the logical and physical worlds, while CIMM operationalizes that foundation as an evaluable schema for LLM-mediated knowledge access.
1. Conceptual role and definition
The study assigns I40 CIM a foundational ontological role rather than the status of a directly queried benchmark artifact. It is the domain model that underlies a community ontology and thereby supplies the classes, properties, and constraints needed for knowledge-graph-based information retrieval in manufacturing. The paper explicitly states that it “focus[es] on the manufacturing domain, particularly on the Bosch Line Information System KG and the I40 Core Information Model,” but the actual evaluation is carried out on two ontologies: LIS and CIMM (Monka et al., 30 Jul 2025).
Within this framing, I40 CIM is the conceptual basis for a schema that supports natural-language-to-SPARQL generation. The practical problem is that manufacturing KGs are accessible through formal query languages, whereas many information needs are formulated in ordinary domain language. I40 CIM, via CIMM, provides the schema-level context that an LLM requires in order to map manufacturing questions to valid ontology terms and relations. The paper’s description makes this function precise: the model defines the structured domain context needed to generate SPARQL over real manufacturing data.
This suggests that I40 CIM is treated less as a standalone executable ontology interface and more as a reference model whose value becomes visible when downstream ontologies inherit and expose its conceptual organization. A plausible implication is that the usefulness of I40 CIM in LLM-mediated querying depends not only on its conceptual scope but also on how that scope is rendered in the accessible schema presented to the model.
2. CIMM as the operational extension of I40 CIM
The ontology actually used to represent the I40 CIM perspective in the experiments is CIMM. The paper describes CIMM as “a community endeavor aiming to describe common concepts in manufacturing,” as a “domain ontology,” as “based on the international IEC manufacturing standard,” and as something that “further extends the semantic expressiveness present in the LIS ontology” (Monka et al., 30 Jul 2025). It is also aligned with I40 CIM and serves as its richer extension in the study.
This distinction between I40 CIM and CIMM is central. I40 CIM supplies the canonical conceptual background; CIMM is the concrete schema used in prompting, ablation, and evaluation. By contrast, LIS is characterized as a production KG and use-case ontology “modeled according to the relational DB schema of the source data.” The comparison between LIS and CIMM therefore contrasts a Bosch-specific schema with a more expressive, community-aligned ontology grounded in the IEC-oriented and I40 CIM-oriented modeling tradition.
The paper uses this dual-ontology setup to test whether context-aware prompting remains effective when the schema is not merely use-case specific but semantically richer and standardized. CIMM is therefore not only an evaluation ontology but also the vehicle through which the paper studies I40 CIM in practice. Its repository is reported as https://github.com/eclipse-esmf/esmf-manufacturing-information-model.
3. Ontological structure, logical–physical separation, and modeling nuances
A defining structural property of the CIMM ontology, inherited from the standard-oriented modeling basis associated with I40 CIM, is the separation of the physical and logical world into different conceptual models. The paper states that “the CIMM ontology … is built based on an international standard that models the physical and the logical world in different conceptual models” (Monka et al., 30 Jul 2025). This split is not presented as an abstract design choice only; it directly affects query interpretation and query construction.
The paper highlights a specific modeling nuance through the class cimm:Equipment. It states that “the concept cimm:Equipment in the CIMM ontology is used to describe logical concepts, e.g., Plant, Lines, Stations.” At the same time, it notes that domain experts often use the term “equipment” to refer to physical assets such as machines. This creates a semantic mismatch between user language and ontology language. In practical terms, a natural-language question about “machines in a plant” may be mapped incorrectly if the model is not given enough schema context to distinguish the ontology’s logical usage from everyday manufacturing terminology.
The scope of the model in the study includes organizational and logical entities such as plants, lines, and stations, their relations, and their mappings to physical-world notions. The paper does not print concrete RDF, OWL, or SHACL fragments for CIMM, nor does it provide explicit triple patterns. It does, however, specify the schema information retained when ontology subsets are prepared for prompting. For reduced structural variants, the retained predicates are rdf:type, rdfs:label, rdfs:domain, rdfs:range, rdfs:subClassOf, and rdfs:subPropertyOf. For richer variants, rdfs:comment, owl:inverseOf, and related axioms are additionally included.
| Variant | Selection principle | Retained information |
|---|---|---|
| Ont_A | Naive reduction of entire ontology | rdf:type, rdfs:label, rdfs:domain, rdfs:range, rdfs:subClassOf, rdfs:subPropertyOf |
| Ont_B | Context-based reduction on Ont_A | Top-25 matching classes/properties plus neighbors |
| Ont_C | Context-based reduction on full ontology | Complete definitions such as rdfs:comment and owl:inverseOf |
| Ont_D | Ont_C plus ontology-based enrichment | Extra generated descriptions for classes/properties |
The paper uses the prefix cimm: but does not provide base URIs in the text. That omission is consequential for reproducibility at the level of explicit namespace declarations, but it does not affect the paper’s focus, which is prompt conditioning rather than ontology publication format.
4. Use in natural-language-to-SPARQL mediation
The study situates I40 CIM, through CIMM, within a pipeline in which LLMs act as mediators between natural-language questions and SPARQL queries over manufacturing KGs (Monka et al., 30 Jul 2025). The key technical issue is context provision: the ontology is too large and too semantically nuanced to be inserted wholesale in all cases, yet insufficient context leads to hallucinated or semantically incorrect query terms.
The paper evaluates several context-aware prompting strategies. Ont_A performs a naive reduction of the entire ontology to structural predicates that fit token budgets. Ont_B introduces question-specific retrieval by selecting the top-25 matching classes and properties via vector similarity and adding neighbors. Ont_C keeps this contextual selection but includes richer semantics such as comments and inverse properties. Ont_D adds heuristic ontology-based enrichment. These variants are passed to the LLM in one of three representations: a Turtle graph, a table listing classes and object/datatype properties, or a table-sorted format in which properties are grouped by class.
The prompts are likewise varied. P_simple uses the instruction: “Write a SPARQL query … Use the following ontology as schema: {ontology}.” P_example adds a generic example SPARQL query. P_domain adds a domain-specific example tailored to the ontology, such as “to find all Materials … used on Line Y in Plant X: {query}.” The study characterizes “adequate context” as a combination of restricted question-relevant schema scope, labels and domain/range information, and, where possible, descriptions and axioms, together with one domain-specific example query.
Representative manufacturing questions include “How many Stations are on Line X from Plant Y?”, “How many equipments are on Line Y in Plant X?”, “Give me the list of material numbers for the line X?”, “For process X: which other lines share that process within the same traceability database?”, and “For production line X: where and how can I access the transactional data for data analytics?” The paper does not print the actual SPARQL queries for these questions, so no concrete code listing is available.
A plausible implication is that the paper treats ontology prompting not as simple serialization of schema metadata but as a controlled schema-selection problem. In that reading, I40 CIM-derived structure becomes useful to the model only after it is filtered into a compact sub-ontology that matches the semantic locus of the question.
5. Quantitative and qualitative findings on CIMM
The benchmark contains 17 business questions drawn from real personas: Benchmarking Engineer, OEM Benchmarking Planner, Maintenance Planner, Technology Developer, and Data Engineer. Each question has a gold SPARQL for both ontologies, LIS and CIMM. The evaluated LLMs are GPT‑3.5‑16K and GPT‑4‑32K. All combinations of ontology selection variant, representation, and prompt template are tested, and each experiment is repeated 5 times with mean results reported (Monka et al., 30 Jul 2025).
The primary quantitative metric is hallucination accuracy over ontology terms appearing in generated SPARQL. The paper defines:
For CIMM using GPT‑3.5, the reported hallucination accuracies are as follows. Under P_simple(graph), Ont_B reaches 0.92, Ont_C 0.63, and Ont_D 0.71. Under P_simple(table), Ont_A yields 0.76, Ont_B 0.87, Ont_C 0.90, and Ont_D 0.84. Under P_simple(table-sorted), Ont_A gives 0.81, Ont_B 0.92, Ont_C 0.92, and Ont_D 0.90. Under P_example(graph), the values are Ont_B 0.94, Ont_C 0.69, and Ont_D 0.75. Under P_example(table), Ont_A is 0.84, Ont_B 0.93, Ont_C 0.95, and Ont_D 0.93. Under P_example(table-sorted), Ont_A is 0.87, Ont_B 0.91, Ont_C 0.94, and Ont_D 0.88. Under P_domain(graph), Ont_B reaches 0.95, Ont_C 0.88, and Ont_D 0.81. Under P_domain(table), Ont_A is 0.85, Ont_B 0.96, Ont_C 0.92, and Ont_D 0.93. Under P_domain(table-sorted), Ont_A is 0.91, Ont_B 0.91, Ont_C 0.92, and Ont_D 0.89.
The paper also reports a GPT‑3.5 versus GPT‑4 comparison on CIMM for graph-based settings. Under P_example(graph), GPT‑3.5 records Ont_B 0.94, Ont_C 0.69, and Ont_D 0.75, whereas GPT‑4 records Ont_A 0.90, Ont_B 0.95, Ont_C 0.85, and Ont_D 0.90. Under P_domain(graph), GPT‑3.5 records Ont_B 0.95, Ont_C 0.88, and Ont_D 0.81, whereas GPT‑4 records Ont_A 0.93, Ont_B 0.95, Ont_C 0.93, and Ont_D 0.91.
The qualitative evaluation uses GPT‑4‑32K and expert judgment on correctness and completeness, each rated on a 0–4 scale. The main findings are that “both correctness and completeness increase when reducing the ontology to a contextualized sub-ontology that only includes the relevant concepts,” that the improvement from Ont_B to Ont_C shows that “the LLMs can leverage the additional semantics provided by the rich ontology,” and that “there exists no significant gain measurable as a result of content enrichment (Ont_D)” because CIMM already contains rich definitions. Aggregated improvements are reported as 23.2% and 29.7% in correctness and completeness, respectively, for Ont_A versus Ont_C, and 13.9% and 22.3% for Ont_B versus Ont_C.
These results indicate that the most important variable is not mere ontology availability but the adequacy of the selected ontology context. The paper also states that adding one domain-specific example in the prompt yields an additional 5–8% accuracy improvement across benchmarks.
6. Error sources, operational guidance, and limitations
The paper’s error analysis on CIMM centers on semantic misalignment, hallucinated ontology terms, and sensitivity to representation choices (Monka et al., 30 Jul 2025). The recurring semantic issue is the mismatch between user wording and ontology modeling. Because cimm:Equipment denotes logical concepts such as plants, lines, and stations, while domain experts often use “equipment” to mean physical machines, an LLM may choose incorrect classes or properties unless the prompt includes enough schema-level clarification.
Hallucinated URIs and ontology terms are directly measured by the hallucination metric. The paper reports that with broader context, especially the naive Ont_A setting, GPT‑3.5 is more prone to such errors. Context-based reduction substantially lowers this risk on CIMM. One explicit example is P_example(table), where the score rises from 0.84 with Ont_A to 0.95 with Ont_C. At the same time, the paper notes that representation alone does not resolve hallucinations: graph, table, and table-sorted forms do not exhibit a universally dominant pattern on CIMM.
From these findings, the paper extracts several operational guidelines. The first is scope control through contextual sub-ontologies, using top‑ semantic similarity retrieval with neighbors; the paper used . The second is to include rich ontology metadata, especially rdfs:comment and axioms such as owl:inverseOf, when available. The third is to provide one domain-specific SPARQL example in the prompt. The fourth is model selection: GPT‑4 outperformed GPT‑3.5 on CIMM under comparable settings. The fifth is post-validation, namely automatic comparison of generated SPARQL subject, predicate, and object tokens against the ontology vocabulary before execution. For critical queries, the paper further recommends qualitative expert review.
The limitations are explicit. LLM non-determinism complicates consistent SPARQL generation, and the authors point to constraints or compensatory mechanisms such as ensemble generation and answer rating. They did not fine-tune because of the scarcity of natural-language-to-SPARQL pairs, making dataset construction for I40 CIM/CIMM-aligned querying an evident next step. The hallucination accuracy metric itself does not capture full execution correctness or completeness, so richer and execution-based evaluation is needed. Finally, the paper emphasizes that CIMM, as an I40 CIM-derived model with separate physical and logical domains, can be difficult even for human experts to navigate when writing correct SPARQL. Future work is therefore directed toward more complex multi-constraint questions, such as capability-based line selection, and toward open LLMs.
In this usage context, I40 CIM is best understood as a manufacturing reference model whose practical significance emerges through an extension ontology, CIMM, and through schema-selection techniques that make its semantics tractable for LLM-driven query synthesis. The evidence reported in the study supports a narrow but technically important conclusion: for I40 CIM-aligned manufacturing KGs, query quality depends primarily on whether the model receives the right ontology fragment with sufficiently rich semantic annotations, rather than on exposure to the entire ontology.