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

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The Transport Research and Innovation Monitoring and Information System (TRIMIS) tracks EU-funded transport research and innovation projects to identify trends and inform policymaking. A key function of TRIMIS is maintaining a comprehensive and accurate database of transport projects. Selecting and categorising relevant projects is a manual and time-consuming process. This study explores the use of artificial intelligence (AI), specifically large language models (LLMs), to support and accelerate this task. A workflow was developed using an LLM platform (GENESIS, the JRC internal, secure, general-purpose LLM platform), iteratively refining prompts to match expert human judgment in identifying relevant projects from CORDIS. Comparing LLM-generated outputs to the existing TRIMIS database, projects were assessed for inclusion/exclusion and classified by transport mode and sector, and technology thematic area. The classifications allow for filtering and consequent analysis and visualisation by these categories. A further application investigated was the assessment of projects for their contribution to the competitiveness of the EU Automotive Industry. Prompt engineering iterations and structured output analysis improved accuracy and consistency and reduced the risk of unsupported or fabricated classifications by requiring every positive decision to be grounded in a verbatim quotation from the project text. Using the ensemble configuration, we analysed 52,923 Horizon 2020 and Horizon Europe projects (2014–2026), identifying 6,986 relevant for inclusion in the TRIMIS database, with reporting in this paper focused on the most recent 19,025 Horizon Europe projects, of which 2,184 (11.5%) were assessed as relevant. The analysis of the full corpus required about two days of serial model time for the two deployed configurations (Annex 5), reducing manual work and identifying matches in previously overlooked funding sources. Outcomes also helped to re-evaluate existing taxonomies. The results demonstrate the potential of AI-assisted workflows to reduce manual workload, improve accuracy, and scale project curation for the TRIMIS database, with broader implications for using AI-assisted methods in public-sector knowledge management. Future work will explore hybrid human-AI review pipelines and domain-specific fine-tuning.