Global Marine Isochore Estimates Using Machine Learning

The thickness normal to deposition (isopachs) and vertical thickness (isochores) of geological units is important for assessing various geologic processes. We present the first marine global sediment isochore estimates for five geological periods dating from middle Miocene (15.97 Ma) to present. We use sparsely distributed sediment depth vs. age observations from the Deep Sea Drilling Project and global maps of biological, oceanographic, geographic, and geological variables as training features in a k‐nearest neighbor regressor to estimate isochores. Results are compared to isochore estimates generated by applying a constant depositional rate from recent estimates of global total sediment thicknesses. Both models of isochore thickness exhibit consistent error. Results from a machine learning approach show major advantages, including results that are quantitative, easily updatable, and accompanied with uncertainty estimation. Final predictions can provide first‐order constraints on sediment deposition with geologic time, which is of timely importance for assessing past climate variability. Plain Language Summary Maps displaying the thickness of sedimentary units are useful for a variety of reasons such as assessing how Earth's climate and depositional systems (e.g., river deltas) have changed through geological time. In this study, we use a machine learning approach to estimate how thick time bound (present day to 15.97 million years) geological units are. To produce these global thickness estimates, we used depth‐age observations from an international deep‐sea drilling collaboration, the Deep Sea Drilling Project. Our machine learning approach uses these observations and complementary data sets, such as water depth and latitude, to estimate unit thicknesses where previous observation has not been made. We then compare the machine learning approach to an independently empirically based method. The model skill calculated by comparing observed and predicted values for each model is approximately the same; however, there are major advantages of using a machine learning approach. Major advantages of using a machine learning approach include the ability to easily update final predictions with new data and calculated uncertainty estimates in the final prediction. Predictions of unit thicknesses can help to indicate where sediment is more likely to accumulate and be preserved in geologic time. Since sediment is capable of sequestering large amounts of carbon from