Cracks are accounted as the most destructive discontinuity in the rock, soil, and concrete. Enhancing our knowledge from their properties is crucial in geo-systems. The most well-known mechanical parameter for such an evaluation is wave velocity. In small scales, such information is obtained using the ultrasonic pulse velocity (UPV) technique as a non-destructive test. In this paper, we take advantage of the recent advancements in machine learning (ML) for analyzing wave signals and predict rock properties such as crack density (CD)‒ the number of cracks per unit volume. To this end, we designed numerical models with different CDs, different wavelength to crack length ratio and, using the rotated staggered finite-difference grid (RSG) technique, simulated wave propagation. Then, convolutional neural networks were used for crack density prediction. Results revealed that wavelength to crack length ratio of 34 leads to best predictions. accuracy of the transmitted signals is slightly higher than the reflected signals. Also, accuracy of 2D signals is marginally higher than 1D signals. However, predations from 1D signals (either transmitted or reflected) are accurate enough (R2 > 97%) that be used for crack density predictions via a machine learning approach.