AI-based knowledge management system for risk assessment and root cause analysis in semiconductor industry

Abstract

Due to the increasing technical complexity of products and market pressure, the demands in the semiconductor industry are rising with respect to quality, performance, and time to market. Root cause analysis and risk assessment are crucial elements for success in fulfilling these demands. As a result, there is an ever-growing number of technical documents, which potentially contain valuable information serving as a base to inform development and production. Experts need to cope with this large number of technical documents, for example, to generate new hypotheses to identify possible root causes of deviations or potential risks in the ramp-up and production phase of new products. Unfortunately, most of the technical documents are unstructured, making processing them even more tedious. New advances in computer science, specifically artificial intelligence (AI), open the door for a higher degree of automation of knowledge management tools to support experts. Knowledge bases such as knowledge graphs allow for representing complex information but need to be created for each domain. Novel state-of-the-art graph embedding algorithms showed promising results in complementing knowledge bases with new relations. Complementary to knowledge base completion, language models trained on large textual corpora have demonstrated their ability to capture complex semantics. This paper proposes a new expert system concept for failure root cause analysis and risk assessment in the semiconductor industry, which leverages the advanced graph embeddings in combination with language models. The main challenges in this setting are the type of relations of interest, which are causal, and the language being used, which is highly domain-specific. Thus, we devised AI for consistency improvement of the data, predicting new links, and information extraction from unstructured data. The information extraction is conducted by levaraging domain specific ontologies and by focusing on presence of causal language.