DETECTive: Machine Learning-driven Automatic Test Pattern Prediction for Faults in Digital Circuits
Vincenzo Petrolo, Sourav Medya, Mariagrazia Graziano, Debjit Pal · 2024
Due to the continuous technology scaling and the ever-increasing complexity and size of the hardware designs, manufacturing defects have become a key obstacle in meeting end-user demand. Despite decades of research, traditional test-generation techniques often struggle to scale to massive and complex designs. Such scalability issues stem from the numerous backtracking the traditional test generation techniques perform before converging to a test pattern. In this work, we present DETECTive that leverages deep learning on graphs to learn fault characteristics and predict test pattern(s) to expose faults without requiring backtracking. DETECTive is trained on small circuits, and its learned knowledge is transferable to predict test patterns for circuits that contain up to 29 × more gates than the training circuits. Since DETECTive avoids backtracking completely, it can predict test patterns up to 15 × faster than academic tools and up to 2 × faster than commercial tools. DETECTive achieves up to 100% pattern accuracy on synthetic designs and up to 95% test pattern accuracy on realistic designs. To our knowledge, DETECTive is the first to leverage deep learning to predict test patterns for digital hardware designs that can complement the traditional test generation techniques for faster design closure.