DETECTive: Machine Learning driven Automatic Test Pattern Prediction for Faults in Digital Circuits

Design for testing is becoming a critical stage of the design process. This procedure helps detect electrical faults on a digital circuit and it is of paramount importance in safety-critical applications. Automatic Test Pattern Generation is the traditional algorithmic approach capable of finding all the test pattern sequences that detect the presence of electrical faults. Unfortunately, due to its NP-Complete nature the decision problem requires backtracking before converging to a solution. Even though heuristics have been developed to decrease the number of backtracks this remains the bottleneck when dealing with industrial-scale designs. To address this problem, we introduce the concept of “Automatic Test Pattern Prediction” (ATPP) that leverages the power of deep learning to predict test patterns instead of generating them. To this end, we present DETECTive, the first fully machine learning-based ATPG tool. The model is trained on small-scale circuits and our findings prove it to be effective on circuits up to four times larger, demonstrating that the model learns how to predict test patterns. Moreover, due to its non-backtracking nature, the model runs 11x faster than academic tools. We think that this research settles the basis for more sophisticated models that can perform predictions on industrial-scale designs. This could lead to a lower time required to reach high fault coverage, a shorter time-to-market for new chips, and safer devices.