A Comprehensive Evaluation of Machine Learning on Coral Trace Element Paleothermometers for Sea Surface Temperature Reconstruction

Abstract This research introduces a novel approach to reconstruct sea surface temperature (SST) by developing a universal coral thermometer using machine learning (ML) algorithms on monthly resolved Porites coral proxies and SST data. A total of 1,202 data sets from 19 corals, covering SSTs ranging from 21.5 to 31.5°C, with proxies including Sr/Ca, Mg/Ca, Li/Mg, U/Ca, and B/Ca ratios were analyzed. The data were divided into four sub‐datasets by regional and taxon constraints. An exhaustive analysis was conducted, training 1,612 models using various proxy combinations and ML strategies to assess the impact of the non‐SST effect on the universality of ML models. The results indicated that the non‐SST effect is more significantly attributed to regional variations than to taxon differences, underscoring the importance of regional factors in Porites coral proxy‐based SST reconstructions. Sr/Ca and Li/Mg proxies were identified as the most indicative of SST, showing clearer relationships with temperature than other proxies. Non‐linear approaches achieved a Root Mean Square Error (RMSE) of less than 0.90°C, which further decreased to 0.72°C upon incorporating specific regional and taxon constraints. In an independent test set focusing exclusively on Li/Mg and Sr/Ca proxies, the tree‐based algorithms particularly excelled, achieving an average RMSE improvement of at least 0.52°C over the Universal Multi‐Trace Element Calibration Scheme and the Li/Mg empirical equation. This research underscores the potential of applying ML to coral‐based SST reconstructions, especially highlighting the effectiveness of tree‐based algorithms and the suitability of Sr/Ca and Li/Mg proxies for accurate temperature estimations.