USING NEURAL NETWORKS TO MODEL SEA SURFACE TEMPERATURES

Quantifying past global temperatures and their evolution is critical for constraining the sensitivity of Earth’s climate system to changes in atmospheric carbon dioxide, and predicting the magnitude of future warming. Understanding how regional sea surface temperatures (SSTs) have changed is also important for reconstructing patterns of oceanographic variability with implications for global atmospheric circulation, monsoonal circulation, hurricane genesis and more. For decades simple transfer functions have been used to relate counts of planktonic foraminifera to SSTs, and to reconstruct past spatial and temporal variations in ocean temperature (see the work of CLIMAP and MARGO).

Here, we use a commercially-available machine-learning architecture generator, the NeuroGenetic Optimizer 3.6, to train and test regionally- and seasonally-resolved models for SST prediction. Our work parallels that of Kucera et al., 2005, but employs a narrower range of model architectures and is more regionally targeted. We use our results to investigate model performance in the tropical Atlantic, to compare our model results with geochemical reconstructions of core top SSTs using the Mg/Ca and U^k’37 geochemical proxies, and to predict Last Glacial Maximum SSTs. We also share progress in developing new open source PyTorch code to replicate the NGO’s artificial neural network design, with the goal of increasing the accessibility of neural network modeling techniques to the paleoclimate community.