Integrating the Theory of Inventive Problem Solving with Large Language Models: Enhancing Reasoning for Innovation in Materials Science at the Molecular Scale

This work proposes a general conceptual framework for integrating TRIZ (Theory of Inventive Problem Solving) structured reasoning into large language model (LLM)-based workflows for chemical and materials science. We argue that persistent AI challenges in this domain—data scarcity, weak scaffold transferability, unphysical predictions, and limited interpretability—are most naturally framed as TRIZ-style contradictions and that embedding contradiction-resolution logic into LLM reasoning can elevate AI from pattern-matching to inventive, researcher-like problem solving. Unlike prior AI–TRIZ integrations such as AutoTRIZ and TRIZ-GPT, which address general engineering tasks, the present framework extends TRIZ tools to physicochemical phenomena and targets local, privacy-preserving deployment. To illustrate the concept and identify directions for further development, we implement and evaluate a simplified three-stage proof-of-concept pipeline on nine local LLMs across eleven chemical problems. Results show that the TRIZ-guided pipeline substantially reduces token consumption—both overall and especially in the solution-generation stage—without an obvious loss in solution quality under the adopted evaluation criteria, suggesting considerable room for further improvement as the framework matures.