Earthquakes represent one of our greatest natural hazards. Even a modest improvement in the ability to forecast devastating events like the 2016 sequence that destroyed the villages of Amatrice and Norcia, Italy would save thousands of lives and billions of euros. Current efforts to forecast earthquakes are hampered by a lack of reliable lab or field observations. Moreover, even when changes in rock properties prior to failure (precursors) have been found, we have not known enough about the physics to rationally extrapolate lab results to tectonic faults and account for tectonic history, local plate motion, hydrogeology, or the local P/T/chemical environment. However, recent advances show: 1) clear and consistent precursors prior to earthquake-like failure in the lab and 2) that lab earthquakes can be predicted using machine learning (ML). These works show that stick-slip failure events –the lab equivalent of earthquakes– are preceded by a cascade of micro-failure events that radiate elastic energy in a manner that foretells catastrophic failure. Remarkably, ML predicts the failure time and in some cases the magnitude of lab earthquakes. Here, I propose to connect these results with field observations and use ML to search for earthquake precursors and build predictive models for tectonic faulting. This proposal will support acquisition and analysis of seismic and geodetic data and construction of new lab equipment to unravel earthquake physics, precursors and forecasts. I will use my background in earthquake source theory, ML, fault rheology, and geodesy to address the physics of earthquake precursors, the conditions under which they can be observed for tectonic faults and the extent to which ML can forecast the spectrum of fault slip modes. My multidisciplinary team will train the next generation of researchers in earthquake science and foster a new level of broad community collaboration.