Abstract The geochemistry of deep‐sea sediments is crucial for reconstructing past climate variations in detail. A range of models has been employed to enhance the resolution of geochemical measurements derived from rapid X‐ray fluorescence (XRF) scanning data. However, the applicability of these conventional models is generally limited to individual research projects. In this study, we propose a transferable, two‐step deep learning approach designed to overcome project‐specific limitations and continually optimize model performance using diverse data sources, including legacy sediment cores. First, we pretrain a self‐supervised foundation model, Masked Autoencoders for XRF (MAX), solely on XRF spectra from 6 scientific drilling projects enabling the model to acquire general knowledge of XRF. Next, we fine‐tune MAX with a small amount of paired data to quantify XRF spectra into two key but costly deep‐sea sediment measurements: and total organic carbon contents (wt%). Our results indicate that MAX, requiring only one‐third of the training data, surpasses conventional models in quantification accuracy. Furthermore, the model's generalizability improves by 20% in zero‐shot tests on new materials with its explainability further supporting its robustness. These advantages suggest that MAX could inspire future development of models with greater complexity and broader data coverage. Moreover, MAX holds the potential for application in other geological and X‐ray‐based fields extending its impact beyond deep‐sea sediment geochemistry and XRF.
Lee et al. (Thu,) studied this question.