New global lacustrine brGDGTs temperature calibrations based on machine learning

Branched glycerol dialkyl glycerol tetraethers (brGDGTs) are ubiquitous bacterial membrane lipids that serve as a reliable proxy for paleotemperature reconstruction. The salinity and seasonal icing of lakes play a crucial role in brGDGT calibration, with calibrations varying between tropical, mid-high latitude, and freshwater versus non-freshwater lacustrine sediments, which complicates their universality. Furthermore, changes in a lake's salinity and seasonal icing during glacial and interglacial cycles can influence the accuracy of paleotemperature reconstructions across different time periods. In this study, we present improved global calibrations based on a comprehensive lacustrine brGDGT dataset consisting of 949 samples, utilizing methods such as linear regression for the fractional abundance of key brGDGTs and support vector machine regression with a linear kernel. The calibrations yield root mean square errors ranging from approximately 2.0 °C–2.6 °C for the test dataset, closely matching those obtained from the training dataset, indicating that the results are not overfitted. These calibrations show improved performance when applied to lake sediment sequences. • Present improved global calibrations using a comprehensive lacustrine brGDGT dataset and machine learning. • Present calibrations for both saline and freshwater lakes. • Demonstrate MAF as the optimal metric for global lacustrine temperature reconstructions by brGDGTs.