Diatom‐Based Climate Reconstruction and Modeling: Modern Calibration Set and Paleodata

Abstract The past 2000 years have been crucial to evaluating climate models and their projections. We explored an artificial neural network (ANN) to reconstruct (the past 600 years) and model the climate for the coming decades. For reconstruction of air temperature (AT) and sea surface temperature (SST), time series of instrumental data and diatoms fluxes, contemporaneous and preserved in a laminated sediment core from the Alfonso Basin, were used to train a feedforward ANN (6 neurons for AT and 14, 14 for SST). We demonstrated that ANNs have an advantage in simplifying and determining the nonlinear relations of the climate system with the assemblages of planktonic diatoms identified as AT and SST proxies. During validation, significant distance correlations were achieved for reconstructions (r 2 dist = 0.65 for AT and 0.63 for SST; p < 0.005) and prediction (r 2 dist = 0.76 for AT and 0.70 for SST; p < 0.005). In this way, a better performance was obtained [AT (SST): r = 0.98 (0.99), Root mean square error (RMSE) = 1.69 (1.88)] than the reconstructions by other authors for the region using conventional methods. Finally, present‐time and paleodata integration into the ANNs for prediction allowed improving the number of steps into the future with greater certainty, precision, performance [AT (SST): r = 0.94 (0.94), RMSE = 2.83 (5.98), Mean Absolute Percent Error = 2.98 (3.97)], and prediction accuracy (>90%), with values within the range of variation of the instrumental data. However, we recognize that increasing data in the ANN training base and higher resolution paleodata from laminated sediments are still necessary, so it is crucial to maintain the regional observatories.