Due to statistical gate delays, the critical probability of a testable path varies among its test patterns. Thus, the delay defect coverage of a selected set of critical paths depends on the selected test set. A new framework to select a test set for improved delay defect coverage is presented. It uses an algorithm that computes the critical probability of a path by a test pattern and machine learning to identify a small test set that maximizes the combined delay defect coverage. Experimental results demonstrate a significant improvement in delay defect coverage over existing static critical path approach.