Knowledge organisation and intelligent R&D for chemicals and materials industries – A workshop report

Page 1 of 15 Knowledge organisation and intelligent R&D for chemicals and materials industries – A workshop report Gerhard Goldbeck1,2*, Vikki Cantrill1, Ferry Kienberger3, and Alexandra Simperler2 1 Goldbeck Consulting Ltd, UK 2 EMMC ASBL - European Materials Modelling Council, Belgium 3 Keysight, Austria * Corresponding Author: [email protected] Page 2 of 15 Contents Contents .................................................................................................................................................. 2 Preface .................................................................................................................................................... 3 Introduction ............................................................................................................................................ 3 Industry perspectives .............................................................................................................................. 3 Lessons learnt from collaborative research projects .............................................................................. 6 More semantics, AI, and EMMO ............................................................................................................. 7 Looking to the future of intelligent R&D ................................................................................................ 8 Key points ................................................................................................................................................ 8 Semantic data management maturity is low, but substantial advances are planned ........................................ 8 Improve data health and semantics at the data source ..................................................................................... 9 “Digital scientist” interacts with multiple electronic and digital systems .......................................................... 9 Interoperability is the least addressed of the FAIR principles ............................................................................ 9 Development of common semantics in materials industry .............................................................................. 10 Between promise and reality, agentic AI and ontologies ................................................................................. 10 Acknowledgements ............................................................................................................................... 10 Disclaimer.............................................................................................................................................. 10 Workshop contributions ....................................................................................................................... 10 References ............................................................................................................................................ 11 Page 3 of 15 Preface This report provides some reflections on the presentations and discussions that took place during the Semantic Materials Workshop 2025 on 10 November 2025 at Homerton College, Cambridge, UK, with 38 in-person and 35 online registered participants. The workshop was organised by Semantic Materials1, the semantic technologies branch of Goldbeck Consulting Ltd2, in collaboration with European Materials Modelling Council3 (EMMC) and the DigiCell project4,5. This report is an appraisal of the event and not an accurate summary of presentations or discussion contributions. Introduction The workshop followed on from the Semantic Materials Workshop 20246, which took place in April 2024 in London. The objective of this second workshop was to share advances, needs, challenges, and experiences in different fields when using semantic technologies and artificial intelligence (AI) and compare and contrast changes in this area since the first workshop. From a European perspective, use cases, lessons learnt and the foundations of semantic technologies within the chemicals and materials sectors were presented and practical solutions shared. Semantic Materials1 conducted a maturity study7 and the preliminary results show that materials industries are aware of semantic data management methods and plan to or are carrying out pilot projects. Significant shifts and changes in the adoption of AI and knowledge management technologies are expected in the next three to five years as these methods become operational and are rolled out across organisations. It is important to understand the complexity and variety of materials that industry deals with. For example, a simple item, like a palm-sized cogwheel, may be a well-defined alloy yet comprise a vast number of microstructures that can alter its macroscopic behaviour. This calls for multiscale modelling and characterisation8,9 that demands and produces a plethora of data that then often relies on human experts for correct interpretation. Materials science lacks agreed terminologies and controlled vocabularies for semantics10. In contrast, chemistry can draw on the IUPAC Gold Book11 for terminology. To address the issue, Semantic Materials1 brings together experts from philosophy, physics, chemistry and materials sciences that strives to put the field on a sound ontological footing and provide knowledge translators to benefit industry. Industry perspectives The pharmaceutical industry collects vast amounts of data. To be useful to people and for AI, all data assets must have meaning and provenance with no ambiguity. Then, solutions, such as knowledge graphs, can enable staff to explore particular data spaces. To succeed, FAIR data12–14 is needed from the outset, which relies on data producers trained on data compliance. Page 4 of 15 It is advisable to start by building small data models that describe a specific entity and can be pulled into an application ontology. This, subsequently, serves as reference model, which describes all entities within research and development (R&D) of an organisation. Controlled vocabularies should be built so that data producers label each entity consistently, as well as using alternative/subjectspecific labels and hidden labels15,16, for example the National Cancer Institute Thesaurus17. In combination with Persistent Unique Identifiers (PIDs) 18, this provides a set of metadata that implement FAIR principles in practice and generates a semantic structure for data specialists that are largely invisible to the data users. As a result, data consumers profit from catalogues using Uniform Resource Identifier19 on data assets and metadata that improve interoperability due to data dictionaries based on the above controlled vocabularies and reference models. These help to explain what each data column is and how it is aligned into a data model. This sort of data management forms the majority of a data scientists’ work, so improving interoperability can notably reduce IT costs. R&D scientists’ appreciation of interoperability as an asset beyond their project will further improve data management across the organisation and importantly make deployment of AI across an organisation much more efficient. This process is best described as “putting the knowledge graph into the data”. So, when data scientists build their data asset, they also catalogue it by following an information standard and simultaneously develop a data dictionary. As and when data scientists find terms that are not in the controlled vocabularies, such terms should be highlighted to the respective developer teams. By following these rigorous workflows data assets fit for AI can be produced. It takes considerable time and effort to develop such systems, so their development may also benefit from pre-competitive collaborations20. From a semantics and knowledge management perspective, the materials industry is very broad and complex. For example, several raw materials may undergo several processes/formulations to give a new product that is sold only to become the starting point for the next organisation in the value chain. Consequently, numerous testing and characterisation techniques are used with a plethora of data formats and subject/domain-specific conventions, which hinder the realisation of FAIR12 data. The resulting data is usually not reusable or interoperable, hence extracting knowledge from it can then require considerable time and effort. In an effort to advance digitalisation, companies have introduced a wide range of systems and tools, including report archives; electronic lab notebooks (ELNs)21; laboratory information management systems22; scientific data management systems (that scrape data from testing rigs and analytical equipment into repositories); modelling and simulation tools and related data systems; environmental, health, and safety software; and enterprise resource planning23 systems. Given the complexity of industrial materials R&D, projects can involve many of the above. However, this approach can be at odds with the researcher, who often likes to try numerous ways to solve a problem, which don’t readily fit into the systems provided. For example, if entering each of these experiments into a system triggers several levels of authorisation, researchers may opt for traditional note taking and log only selected successful experiments, without realising the onward impact of the information/knowledge lost. The situation persists because the materials industry does not have the Page 5 of 15 same regulatory requirements as pharma and life sciences areas have that incentivises more standardised ways of knowledge organisation. Yet, the materials industry is well aware of importance of FAIR data and metadata; however, semantics is abstract to many research scientists who lack knowledge to its potential benefits. Interoperability is not considered a priority because the field is very broad and many research areas have been able to thrive without extensive knowledge management. A way forward here could be to encourage people think about the data and information they get from outside/other domains and how that could be incorporated. The experience with AI tools so far is that they are often overpromoted as able to extract information from any data. There is an opportunity for semantic data experts to raise awareness, demystify AI and how it best works in combination with FAIR data. However, more user-friendly tools and interfaces will be required to entice scientists to adopt semantic technologies. An example from the aerospace industry demonstrates the benefit of semantic technologies to support the industry’s sustainability objectives. Although the recycling of metal parts, such as aluminium, is quite successful, composite materials are often incinerated for energy recovery or end up in landfill. New generation aircraft will comprise thermoplastics that are recyclable and can end up again within the sector24,25. To support this, the researchers gathered key data on the materials features and properties, and process parameters along with information from literature, databases, modelling software and lab tests, so that they could identify optimal recycled blends for materials. However, composites are not uniform, so data on the thermoplastic, reinforcement fibres, defects, and so on needs to be included, which adds to the complexity of the problem. This challenge was addressed by using a knowledge graph in conjunction with a bespoke ontology that covered “product” and “process”. In the graph logical model each node was a product and each relationship between nodes a process. Data was extracted automatically with a Named Entity Recognition model to automatically identify and classify entities. Battery development and production is an example of an “everything problem” that requires linked knowledge across many domains to solve it; hence interoperability is essential. All data generated and used must be high quality because end users rely on its accuracy. In this industry sector semantic knowledge organisation is used in individual and hierarchical data models with an emphasis on metadata integration so data can be used across the divisions of an organisation or between scientific disciplines. EU project DigiCell4,5 looks at the characterisation of materials, such as lithium-iron phosphate and lithium-ion nickel, manganese and cobalt batteries. The project aims to test batteries accurately through physical electrochemical models and create digital twins. Such batteries must also follow the international ISO26 and electro technical standardisation27 to be deployed globally and work safely in a variety of devices. Instruments need to be well calibrated so that each measurement is accurate and reliable, so metrology plays an important role, which makes hardware expensive. Therefore, digital twins can be more economical for the battery and for the environment it operates in. Digital twins in the battery sector include characterisation of surfaces, 3D finite element method and 3D scanning electron Page 6 of 15 microscopy, electron microscopy and tomography. This led to the development of the Battery Testing Ontology (BTO)28, which provides a comprehensive framework that represent data and protocols in battery testing and quality control. The BTO is fully aligned with the Elementary Multiperspective Material Ontology (EMMO).29 Sometimes an ontology can help with knowledge organisation and sometimes a large language model (LLM) is sufficient. For example, batteries in electric cars need to be connected to a national power grid to charge. The regulations that need to be followed to do so can be retrieved by a LLM and the information gathered used to automate tests to ensure compliance with regulations in different countries. The interoperability efforts in the battery field have been supported by a range of European research projects that actively develop and use semantic technologies30. This approach can result in a digital twin that acts as a coherent and persistent framework for linking (data) assets. EU project BIG-MAP31,32 and Battery2030+33 developed the Battery Interface Ontology (BattINFO)34, a semantic resource with essential terms and relationships to describe battery cells, materials, methods, and data. BattINFO is an extension of the EMMO29, adheres to World Wide Web Consortium 35 best practices and offers a collection of concepts to describe battery knowledge that build on International Electrotechnical Commission standards27 and the IUPAC Gold Book11. BattINFO34 enables linking of data from and across research data and other prominent sources, such as Wikidata36, DBpedia37 and Pubchem38. A researcher can also retrieve additional information on chemicals from InChI39 keys and CAS registry40 numbers. BattINFO includes Schema.org41 markup for semantic web integration to enable results from the Google Knowledge Graph42 and Google43. Such a foundational resource enables practical implementations that address challenges such as interoperability of timeseries test data from different equipment providers, and the exchange of model parameters between different battery simulation packages. Ultimately, BattINFO is an open-source battery knowledge base that consolidates data sets, metadata and semantic technology to create a coherent and persistent description of battery resources. Lessons learnt from collaborative research projects When working collaboratively on publicly funded projects, it is essential that there is a clear and detailed data management plan that all partners adopt and conform to. The plan should follow the data lifecycle: planning, acquisition, analysis, preservation, sharing, and reuse of data and be sufficiently detailed to ensure it is instructive. Data is often acquired as an Excel or csv file. Consequently, data analysis requires better tooling so workflows and data provenance can be conserved, which are key for reusability and interoperability. Tools, like ELNs, help but effort is needed to understand and record metadata. From the start, data stewards need to assist data producers to create FAIR data. To support that effort, semantic technologies need to be implemented and ready to use for data producers from the start. Page 7 of 15 More semantics, AI, and EMMO The UK is building an infrastructure for physical sciences research data, called the Physical Sciences Data Infrastructure (PSDI)44. The metadata in the PSDI is organised into resource themes that groups data services, tools and guidance that is searchable across all data sources. Metadata45 is crucial and uses one common language to cover all the physical sciences. The infrastructure started very recently and a simple technology stack is available. PSDI takes a top-down approach, starting with metadata about resources [based on the Data Catalogue Vocabulary (DCAT)]46 and then tackles metadata about the data or data provenance metadata. The approach is based on the Cross-Domain Interoperability Framework47 — a set of principles and domain-agnostic standards to support FAIR data interoperability. Agentic AI aims to achieve autonomous, goal-oriented decision-making and actions. Within materials science such a system could identify a new material with desired properties and then generate the necessary code for robots to go ahead and make. This would supersede people writing code on systems that have little or no autonomy. Agentic AI systems work optimally when context is provided, and knowledge graphs are a good way to do this. For example, a multi-agent system could optimise how a task is done by using graphing methods. For R&D, this could create a kind of “autonomous scientist” in which an LLM extracts data from thousands of scientific papers and creates a well-structured knowledge graph. There might be an ontologist agent monitoring nodes and relationships, defining terms or cleaning the ontology. There could be scientific and reviewer agents then deployed to discover a new material. Applying agentic AI to robots requires careful consideration. Once the robot is given a task, it acts. The action must be right the first time or there could be serious consequences. As a safeguard, the coding agent can be provided with a machine-readable “policy card” to encode laws, regulations, ethics and policies. Ontologies can form the foundations for such agentic AI guardrails. Since 2018, a team of scientists have been developing the EMMO29 to build a knowledge framework able to represent the materials world including its physics foundations. Existing top-level ontologies did not satisfy conceptualisations from the applied sciences. For example, quantum phenomena were absent and relativity ignored. So, they developed a multiscale and multiphysics approach to represent entities and their interactions and interrelations, from a very fundamental “micro-physics” level up to a macroscopic level. As a result, EMMO provides a common language that is able to express all aspects of the material world in a way that is consistent with a scientific world view. What are the best practices to implement semantic technologies? We need to start small and gradually build up semantic models and data integration of the processes a client uses and the materials involved. Data is typically logged in different systems; perhaps paper records were digitised and evolved over time. Both processes and systems will have evolved over time, so managing the various data can become complicated. Here, knowledge graphs facilitate linked data and enable clients to query it. However, proof of concept is needed to demonstrate the value of bringing data together in the first place. During implementation, case models are essential to check the validity and ease-of-use of what is being introduced. Easy user interfaces are often preferable to SPARQL48 queries Page 8 of 15 for clients. Once a client understands that semantic technologies can bring efficiencies, next a developer can identify other questions users might want to query data on and expand the model stepwise. In relation to problem solving, semantic technologies can find issues faster so “time-out-ofmarket” can be reduced, the latter being an important business factor. Looking to the future of intelligent R&D Researchers readily engage with AI projects, but less so with semantics or data governance, which are vital for more intelligent R&D. Interest in this fundamental intelligence work needs to be encouraged in such a way that people want to take ownership of it and be actively involved. For example, a data expert identifies a query that is notoriously slow due to lack of interoperability and produces a case study to demonstrate the benefits semantics bring. Semantics also add a level of confidence to an answer because queries can be guided by established facts. LLMs work with text and rely on language to describe the world, whereas sciences use mathematics and data. Here, the epistemological limit of LLMs can hinder scientific advances and interactions without supportive semantics49,50. Scientists also move on to new fields relatively frequently, which would necessitate retraining an LLM. Here, a decent foundational ontology would facilitate moves to new topic areas. Having a common semantic reference is seen as highly valuable in the above contexts. A number of fields including pharmaceuticals and life sciences benefit from many years of joint efforts on ontologies. In contrast, currently, there has been much less incentive for materials and manufacturing to work on common ontologies because the breadth of materials science means limited topic overlap. An exception to this is the battery sector that, owing to a series of EU projects33,51, can use semantics and ontologies for intelligent R&D. Now may be the right moment for materials science field to agree on and establish a common semantic basis, as a number of factors align: the need for data integration across domain, functions and value chains is ever increasing; according to the Gartner hype cycle52, knowledge graphs as a technology sector have entered a mature and more productive state of development; the synergy between AI/LLMs and semantics is well established (ontologies improve AI results and vice versa, AI/LLMs can be used to speed-up ontology development and improve useability), and due to the EMMO and its ecosystem of discipline and domain ontologies — such as the Characterisation Methodology Domain Ontology53 — there is a strong semantic foundation. Key points Semantic data management maturity is low, but substantial advances are planned Preliminary results of the Semantic Materials maturity survey7 are: • Current situation is the maturity is close to level 2: “Active”, meaning there are experimental or pilot projects but there is no or little operational implementation in business functions. • Organisations aim to lift their maturity to between levels 3 and 4 across People, Processes, Data and Tools and Technologies, which means operations with clear objectives and Page 9 of 15 implementation in some functions, working towards systemic, robust and scalable implementations across the business. • The biggest issues organisations aim to address are data from several functional areas that need to be manually collected and integrated; no common definitions of business relevant data and fields used by multiple functional areas; and several “versions” of similar data exist across systems and functional areas. So ultimately, it is unclear what the source of truth is. Improve data health and semantics at the data source Generally, the labelling of experimental data produced is of relatively poor standard. This means the barrier to achieve FAIR data, in particular interoperability, is insurmountable. Efforts to address the issue include: • Ensure data are assigned PIDs. • Catalogue data by using established metadata standards, such as DCAT46. • Establish and use controlled vocabularies and data dictionaries. These efforts can be referred to as “put the knowledge graph into the data” rather than “put the data into the knowledge graph”. Don’t defer proper metadata until the point of data integration, rather work with the data producers to ensure good “data health” to enable linked data. “Digital scientist” interacts with multiple electronic and digital systems The efforts for FAIR data and digitalisation have led to the introduction of a wide range of systems and tools. However, the introductions of more and more systems, each with their own requirements to use them, is not compatible with R&D working practices. Although each system claims to make data FAIR, often they create more metadata siloes, unless there is adequate central management of metadata, dictionaries and ontologies. Interoperability is the least addressed of the FAIR principles The main goal of interoperability principle in FAIR data is to provide a “common understanding” of digital objects by means of a language for knowledge representation to be used to represent these objects54. By definition, interoperability involves different systems and/or functions, so it tends to be overlooked and is the least addressed of the FAIR principles. At the same time, strategic R&D needs interoperability even though existing tools struggle. Welldocumented, shared semantics (controlled vocabularies, metadata schema, ontologies) is key to overcoming interoperability issues. Linked data opens up resources, for example, interoperability between test data from different equipment providers.