LIPOTHERM: A Calibration Framework for GDGT Paleothermometry Based on Machine Learning

Abstract

SST reconstruction from glycerol dialkyl glycerol tetraethers (GDGT) preserved in marine sediments is an important tool in quantitative paleoclimatology. These reconstructions require calibration frameworks that provide mapping between sediment GDGT data and temperature values. Improvements in modeling and data coverage can reduce bias in existing calibrations, such as those based on the univariate TEX₈₆ index and other multivariate frameworks. GDGT cyclization is also known to be affected by nutrient stress as a non-thermal control. In this study, we present LIPOTHERM (Lipid Paleothermometry), a new GDGT-SST calibration trained on an extensive collection of globally distributed core-top GDGT data using supervised machine learning. LIPOTHERM features substantially lower calibration root mean square error than existing forms. Inclusion of core-top nutrient conditions as auxiliary predictors further lowers the calibration error. LIPOTHERM demonstrates that robust paleothermometry frameworks can be constructed by accumulating incremental improvements over different methodological refinements involving algorithm selection, feature space representation, expanding data coverage, and adding auxiliary non-thermal covariates.