Promising results have been attained in the laboratory, where time to failure of simulated gouge in a double-direct shear apparatus was predicted using shallow machine learning methodology, with acoustic emission data as inputs (Rouet-Leduc et al., 2017). Furthermore, earthquake catalog data derived from a double-direct shear apparatus was used to predict timing of upcoming events, shear stress, and friction (Lubbers et al., 2018), showing that simplified representations of raw continuous data sets might hold predictive signals. Machine learning methodology has been applied with great success to the engineering fields, where deep learning algorithms have been used to predict the remaining useful life (time to failure) of mechanical equipment such as roller bearings (Deutsch & He, 2017;Li et al., 2018;Ren et al., 2017). With regard to earthquake prediction, however, time to failure alone is not enough. As a society, we care about predicting the one or two large events that take place in a given year, rather than the tens of thousands of small events that cause little damage. Corbi et al. (2019Corbi et al. ( , 2020) used a subduction zone analog model to address the problem of size and timing prediction. The megathrust interface in their model consisted of two velocity-weakening asperities that ruptured quasi-periodically in response to constant loading boundary conditions. The model experienced a range of earthquake magnitudes, as the fault underwent either complete or partial interface failure. They developed multiple machine learning models, using trench parallel and perpendicular surface displacements as inputs, to predict time to failure (TTF regression;Corbi et al., 2019) and failure imminence (binary classification; Corbi et al., 2020) along nine discrete along-strike zones. Their model was apparently able to recognize precursory displacement associated with the asperities. When both asperities were emitting