Identifying technological problems and exploring potential solutions to support R&D: Tech-mining with multi-LLM applications
Abstract
The shortening of technology lifecycles and growing technological complexity have made the rapid identification of technical problems and exploration of solutions a critical challenge. Patent analysis can serve as a crucial tool for supporting R&D activities by extracting problem–solution pairs from patent documents. However, current patent analysis methods have several limitations. They focus primarily on extracting technical problems from patents without providing a systematic framework for transforming fragmented problems into comprehensive technological intelligence. Moreover, they remain heavily dependent on technical experts, with supervised learning models struggling to achieve the expected performance on unseen data, which is a particularly critical issue given the accelerating technological change. To address these limitations, this paper proposes a systematic framework that converts fragmented technical problems into actionable technological intelligence. The framework performs target–symptom decomposition, inter-problem causal analysis, and solution portfolio development, while employing a multi-large-language-model (multi-LLM) architecture with few-shot learning capabilities to reduce the dependence on technical experts and enhance analytical reliability. A case study of hydrogen storage technology with cross-domain demonstrations in the semiconductor and biopharmaceutical fields demonstrates that the framework can systematically identify technical problems, trace their root causes, and explore potential solutions to support R&D planning with improved cost efficiency and domain-independent applicability.
