Predicting Marine In Situ Heat Flow Using a Geospatial Machine Learning Conformal Prediction

Abstract

AbstractHeat flow measurements over the past decades have drastically increased in number. However, even with vast increases in measurements, large geospatial gaps exist, most notably across Earth's oceans. As marine heat flow data acquisition can be tedious, time consuming, and ultimately expensive, new methods have arisen in an attempt to quantify heat flow across the seafloor using empirical proxies. The goal herein is to obtain a high quality, fully data‐driven, global estimate of marine in situ heat flow with calculated uncertainty. We present a heat flow analysis for Earth's marine environment using a geospatial machine learning algorithm (MLA) utilizing 33,746 heat flow measurements from the International Heat Flow Commission's most recent Global Heat Flow Database. The mean heat flow for the observational data set is 95.83 mWm−2 with a spatially averaged heat loss of 40.7 TW. We use a random forest regression MLA combined with a conformal prediction to predict in situ heat flow and its associated uncertainty for marine environments at a 100 km2 global resolution. Ten‐fold cross‐validation yields an R2 value of 0.4, indicating predictive skill. Geospatial prediction results in a global marine surface weighted in situ heat flow average of 69.67 ± 20.01 mWm−2 with a corresponding heat loss of 29.59 ± 5.78 TW. Additionally, parametric isolation is calculated to highlight areas where additional heat flow measurements would be beneficial to directly improve the overall prediction. The resultant in situ heat flow and parametric isolation maps serve as a tool to further heat flow studies within the community.