Study Evaluates Graph-Augmented Decision Support System for Cleantech Innovation Scouting

In an exploratory design-science evaluation of an evidence-grounded decision support system for cleantech innovation scouting, a hybrid retrieval configuration achieved a Recall@100 of 0.951 and a Precision@10 of 0.278 across an in-corpus benchmark of 50 broad-cleantech queries. However, none of the tested retrieval contrasts reached statistical significance after global Holm correction. The investigation examined system behavior and evaluation sensitivity using relevance judgments derived from two large language model judge pipelines alongside an exploratory case audit.

The evaluated system artifact, designated GRaph-Augmented INnovation Scouting (GRAINS), was designed to address the need for analysts to integrate heterogeneous cleantech evidence while maintaining traceability. The architecture links governed evidence preparation, hybrid retrieval, optional graph expansion, and staged cite-or-abstain synthesis. Its structural contribution couples an inspectable evidence-to-answer pipeline with an explicit rationale connecting engineering choices to scouting requirements.

Analysis of top-10 ranking quality, measured by nDCG@10, revealed that performance outcomes depended on the specific large language model judge pipeline used and whether query-source documents were excluded from the evaluation set. Constraining the number of retrieved chunks allowed per document increased document diversity, though it resulted in a modest reduction in nDCG@10. Furthermore, the evaluated graph-expanded configuration produced lower observed early-ranking scores than the hybrid retrieval baseline.

The evaluation also incorporated a separate exploratory case-based audit of six battery-domain cases assessed by a single researcher. This audit demonstrated how the architecture organizes supported industry actors, applications, and uncertainties, while also identifying existing gaps in task-specific evidence.

The researchers noted that these findings provide bounded, system-oriented design observations regarding the behavior and evaluation of evidence-grounded scouting tools. Because the initial benchmark relied on automated judge pipelines and a single-evaluator audit, the authors concluded that independent human relevance assessments and analyst-workflow studies remain necessary for broader validation.

Subject of Research: Evidence-grounded decision support for cleantech innovation scouting

Article Title: Evidence-grounded decision support for cleantech innovation scouting: a system-oriented design-science study

Article References: Gerber, A. O., Lipenkova, J., Rao, S. X., & Lu, G. (2026). Evidence-grounded decision support for cleantech innovation scouting: a system-oriented design-science study. Frontiers in Artificial Intelligence, 9, Article 1890105. https://doi.org/10.3389/frai.2026.1890105

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1890105

Keywords: cleantech innovation scouting, decision support systems, hybrid information retrieval, graph expansion, knowledge graphs, retrieval-augmented generation, design science, Evidence-grounded, decision, support, cleantech, innovation