Reviving Obsolete Patents: LLMs for Innovation Discovery and Cross-Domain Application

This project addresses the widespread underutilization of patents—where over 95% remain dormant—by designing an AI-assisted workflow that repurposes them as seeds for new ideas. The team sought to systematically uncover overlooked potential in existing IP by using LLMs to identify hidden connections and enable more effective commercialization strategies.

‍

The team developed a five-step framework grounded in prompt engineering, citation tracking, modular decomposition, and cross-domain analogy generation. Starting from a “base patent,” they built out chains of related patents, then extracted reusable technology modules and core problems using GPT-4 and Claude 3.5. These elements served as the building blocks for reframing original problems into new domains using analogy and abstraction.

‍

After extracting technologies and problems into structured data, the team transformed this information into graph-based representations using Neo4j. This allowed them to track innovation lineage, identify white space opportunities, and visually explore how modules and problems connect across domains. These graphs not only improved transparency but also enhanced idea generation through visual reasoning.

‍

The core creative mechanism was analogy generation. LLMs were prompted to suggest new domains where similar problems occur, leveraging their associative reasoning capabilities. For example, a patent on autonomous navigation was analogized to underground utility mapping. This helped uncover surprising but viable applications, supporting exaptation—the reuse of existing technologies in unexpected contexts.

‍

Once new application ideas were proposed, LLMs were further prompted to evaluate them for feasibility, market potential, and technology fit. Market research tools and industry reports like McKinsey Insights were integrated to assess economic viability. This step ensured that the ideas generated weren’t just novel, but aligned with real-world demand and business value.

‍

Looking ahead, the team proposed enhancements including self-evaluating LLMs, retrieval-augmented generation (RAG), and KNN-based prompting to improve relevance and reduce hallucinations. The goal is to build a closed-loop system that continuously learns from past mappings and generates higher-quality recommendations over time. Ultimately, the project shows how LLMs can transform dormant IP into dynamic sources of cross-sector innovation.