Modern semiconductor manufacturing requires accurate fault detection across high-dimensional, multivariate sensor data. Traditional anomaly detection methods often struggle under conditions of noise, sparsity, and limited labels.
This talk introduces the NV-Tesseract-AD within the NV-Tesseract 2.0. model family, a diffusion-based anomaly detection model developed by NVIDIA’s Applied AI Lab. By combining denoising diffusion processes with curriculum learning, the model captures latent patterns of normal system behavior, enabling precise anomaly detection via reconstruction error.
Evaluation of publicly available semiconductor manufacturing datasets reveals significant performance gains when paired with novel adaptive thresholding techniques such as Segmented Confidence Sequences (SCS) and Multi-Scale Adaptive Confidence Segments (MACS). These methods significantly improve F1 scores and robustness over traditional baselines.
The session concludes with an overview of NVIDIA Inference Microservice (NIM) deployment, enabling scalable, real-time integration of NV-Tesseract-AD into semiconductor manufacturing workflows.
