Die-to-prompt: visual language model-based defect inspection and anomaly detection

Currently Die-to-Die(D2D) and Die-to-Database(D2DB) methods for patterned wafer and reticle inspection have several limitations, including expensive and lengthy data collection, highly skewed dataset, long time-to-market and model not robust for different layer or when process drifts overtime. In this work, we’ll demonstrate zero shot defect inspection without needing reference image or model training using the NVIDIA Cosmos Nemotron vision language model (VLM). A prompt and target image could be sufficient to find defects reliability. The same out-of-box VLM can be deployed with NVIDIA video search and summarization (VSS) agent blueprint for anomaly detection in chip manufacturing production.