AI screens 52,923 EU research projects to map transport innovation in days

Curating a database of thousands of European transport research projects has always been slow, painstaking work — a task that analysts at the European Commission’s Joint Research Centre (JRC) have long handled by hand. Now, a study published in European Transport Research Review shows that a carefully engineered artificial intelligence pipeline can do much of that screening in roughly two days of computing time, while flagging relevant projects that human curators had overlooked.

The Transport Research and Innovation Monitoring and Information System, or TRIMIS, tracks EU-funded transport research and innovation projects to identify trends and inform policymaking. Its core function is maintaining a comprehensive, accurate database of transport projects, but selecting and categorising relevant projects from the vast Community Research and Development Information Service (CORDIS) has been a manual, time-consuming process — one that can also lead to inconsistencies, including the potential oversight of relevant projects funded through bottom-up schemes not associated with transport research.

The new study, led by Ilias Cheimariotis and colleagues at the JRC, explores whether large language models (LLMs) could support and accelerate this task. The team built its workflow on GENESIS, the JRC’s internal, secure, general-purpose LLM platform, and benchmarked four open-weight models: GPT-OSS-120B, Llama-3.3-70B, MiniMax-M2 and Mistral-24B. Rather than fine-tuning models on labelled data — which the TRIMIS taxonomy lacks — the researchers relied on prompt engineering, iteratively refining instructions until the model’s judgements matched those of expert human curators.

A key safeguard was built into the prompt itself: every positive relevance decision had to be grounded in a verbatim quotation from the project text. If a supporting quotation could not be located as an exact substring of the input, the decision was detectable and could be rejected. Combined with a fixed JSON output schema and near-deterministic decoding at zero sampling temperature, the researchers say these mechanisms directly constrain the fabrication of evidence and reduce unsupported classifications.

The relevance workflow scored each project on a three-point scale: a score of 2 for projects whose primary subject is transport research or innovation, a score of 1 for projects substantively relevant to transport but where transport is one contributing dimension among others, and a score of 0 for exclusion. Both scores 1 and 2 lead to inclusion in the database. The team tested twelve prompt variants in a factorial design crossing two scoring definitions with six prompt styles, evaluated on a random sample of 250 CORDIS projects against reference labels produced by three TRIMIS analysts who scored independently and agreed unanimously on 85 percent of projects.

On that reference sample, the deployed GPT-OSS-120B configuration reached a macro-F1 of 0.94 on the binary include/exclude decision, with include-class precision of 0.97 and recall of 0.85. The production deployment, however, used an ensemble of two configurations — GPT-OSS-120B and Llama-3.3-70B run in parallel — with projects on which the two disagreed, and where the recall-oriented configuration reported low confidence, routed to a compact expert-review queue. On the 250-project sample, just 13 projects (5.2 percent) were flagged for expert review, a 95 percent reduction in manual screening load.

Applied to the full corpus of 52,923 Horizon 2020 and Horizon Europe projects recorded in CORDIS up to 1 March 2026, the workflow identified 6,986 projects as relevant for inclusion in the TRIMIS database. The paper’s reporting focuses on the most recent 19,025 Horizon Europe projects, of which 2,184 — 11.5 percent — were assessed as relevant: 977 as TRIMIS-core and 1,207 as TRIMIS-extended. The full analysis required about 48 hours of serial model time for the two deployed configurations, or roughly 25 hours if run concurrently — a task the authors note would typically require several person-months of manual work.

Perhaps the most striking finding came when the workflow’s output was cross-examined against the manually curated TRIMIS record for the same period. The AI independently flagged 95.0 percent of the 2,448 curated projects as relevant, reproducing expert judgement without access to the curators’ decisions. But it also surfaced a substantial body of transport-core projects absent from the curated record. Roughly three in five of these candidate omissions sit in non-thematic funding instruments — the European Research Council, Marie Skłodowska-Curie Actions, the European Innovation Council and the SME Instrument — where transport-relevant work is funded without any transport-specific call to filter on, and where call-based manual curation cannot practically reach.

Beyond relevance screening, the team developed classification workflows that assigned projects to transport mode, sector and technology thematic area using an ensemble of two models with complementary dispositions, resolving disagreements through adjudication. Agreement with curated labels was strongest for transport mode, at 70 to 81 percent, and the analysis exposed shortcomings in both the AI’s interpretation and the existing labelling system — including a large residual ‘Other’ thematic class and a signature, shared by about one in five classified projects, that points to candidate taxonomy extensions. The outcomes also helped the team re-evaluate existing taxonomies, since the curated database often places projects under a single category whereas the AI attributed multiple non-mutually-exclusive labels.

As a case study, the researchers applied the same pipeline to assess the contribution of 150 EU automotive research projects to the competitiveness of the European automotive sector, scored across eleven dimensions including consumer price, technical novelty, EU leadership, production scaling, raw materials, circularity and skills development. On the strict reading, explicit EU-leadership evidence — specific first-of-a-kind European capabilities — appeared in roughly a fifth to a third of projects, and genuine manufacturing scale-up in about a sixth, while circularity was the most strongly represented sustainability-related dimension at roughly 31 percent of projects. Actions on consumer price and production scaling were comparatively scarce, pointing to candidate focus areas for future research calls. The authors stress that the assessment is a portfolio-level screening aid validated on 20 double-scored projects, and that broader independent validation would be required before its scores are used for individual-project assessment or formal policy evaluation. Relevance screening and classification, by contrast, are moving toward operational, human-in-the-loop deployment within TRIMIS’s curation process, with experts remaining the authority — specifying dimensions, calibrating definitions, and reviewing the flagged boundary cases.

Subject of Research: Use of large language models to automate curation of the EU TRIMIS transport research database

Article Title: Enhancing transport innovation knowledge management workflows using artificial intelligence: a TRIMIS database case study

Article References: Cheimariotis, I., Neves Costa, J., Piazza, F., Lodi, C., Rataj, M., Zawieska, J., Stepniak, M., & Grosso, M. (2026). Enhancing transport innovation knowledge management workflows using artificial intelligence: a TRIMIS database case study. European Transport Research Review, 18(1), Article 98. https://doi.org/10.1186/s12544-026-00854-w

Image Credits: AI Generated

DOI: 10.1186/s12544-026-00854-w

Keywords: artificial intelligence, large language models, TRIMIS, Horizon Europe, CORDIS, transport research, EU policy, prompt engineering, automotive competitiveness, knowledge management, JRC, human-in-the-loop