Systematic linking of innovation problems with pluridisciplinary solution concepts based on heterogeneous multi-domain data sources using AI and TRIZ

Overview & Synopsis

Patent documents contain a vast record of technical solutions, but their length and domain specific language make cross domain search unreliable. This thesis builds the resources and methods needed for patent based systematic innovation. DAPFAM, a family level benchmark with in domain and out of domain splits, shows out of domain retrieval is roughly five times harder and that dense retrievers lose their edge outside familiar domains. PatenTEB extends evaluation beyond retrieval and supports leakage free embedding comparison, leading to the patembed multi task models that generalize better externally. PatentSpark integrates retrieval, expert curation, LLM synthesis, rarity scoring and end to end provenance. Industrial studies validate the workflow through prototypes.