Chemical vapor deposition (CVD) reactors are used in the process of depositing a layer of a certain substance onto a surface. One of our manufacturing customers operates several such reactors in their Semicon business line.
When such a reactor stops during production, this is called a ‘run-abort’. Leading to scrap at an average cost of over EUR 18k, and lower yield. The raw material is wasted, as is the production time. Before the reactor can be restarted, the root cause of the failure must be found. As there is high demand for Semicon products, lost production time quickly leads to delivery delays and customer complaints.
Quality engineers and process engineers spend significant time to find root causes of aborts to prevent them in the future. They gather data, analyze, create hypotheses, gather more data, analyze again, prepare conclusions, etc. Until they have found the root cause. Manual gathering of data alone takes 4-6 hours per abort.
The complexity of the CVD process means there are many different failure modes that could cause a run-abort. Therefore, there is much data to be gathered from several sources and in different formats. This makes gathering data a tedious and lengthy process.
We have built a secure GenAI assistant, that engineers interact with in natural language and that structures its answers instead of dumping raw data.
Recipe changes, material numbers, logbooks, temperature, and pressure, all relevant data is available in minutes for further analysis just by typing a question to the GenAI assistant.
As LLM, we used Azure OpenAI and added LangChain to provide for easy interaction with the assistant based on natural language. To enable secure access, we used Azure web services combined with Azure Entra ID. And to ensure transparency of choices made by the GenAI assistant, we implemented Langfuse.
With the secure GenAI assistant, engineers now gather data in 5 to 10 minutes, eliminating 25 hours of manual work every week. On top of that, the reactor can restart sooner, which saves valuable production time. And with every abort prevented, 18k worth of raw material is saved.