Résumé
Abstract Branched glycerol dialkyl glycerol tetraethers (brGDGTs), bacterial membrane lipids, are widely used as temperature proxies. Although these proxies are effective for reconstructing past temperatures, brGDGT‐based methods encounter limitation in extreme climate gradients and confounding factors such as salinity and aridity, conditions particularly prevalent in drylands. In Arid Central Asia (ACA), the commonly used linear brGDGT calibration exhibits significant biases, resulting in substantial errors in temperature reconstructions. This study compares two machine learning regression methods, Random Forest (RF) and Boosted Regression Trees (BRT), with traditional linear calibrations, using 761 surface samples from the ACA surface database. It also evaluates an unsupervised machine learning approach based on cluster and weighted combined calibrations. The analyses focuses on the robustness of BRT and RF methods, their effectiveness to reduce biases specific to drylands, and their performance in reconstructing Holocene climate, especially temperature and aridity. The results demonstrate that (a) machine learning methods reduce biases caused by confounding factors in drylands and improve temperature accuracy compared to linear models; (b) among the tested approaches, BRT outperforms RF in climate reconstructions; (c) machine learning enables independent and reliable predictions of both temperature and moisture; and (d) cluster‐based calibrations provide additional improvements in specific archive contexts. This new machine leaning‐based framework increases the reliability of brGDGT‐derived climate reconstructions in drylands. Nonetheless, challenges persist due to past environmental fluctuations affecting the study context. Further expansion of surface data sets across a wider range of climates and archive types is essential to strengthen machine learning training and improve model precision.