Identifying and correcting climate projection biases using artificial intelligence

Abstract:

The fidelity of climate projection, which is crucial for inferring climate change and informing adaptation decisions, is often impaired by biases in climate models due to their simplification or misrepresentation of the climate system. While various bias correction methods have been developed to post-process model outputs to match observations, existing approaches often focus on limited, lower-order statistics, break the consistency between model dynamics and diagnostic variables, fail to generalize beyond the calibration period, offer little implication for model improvements, even conceal crucial model biases. Here we introduce a novel bias identification and correction method that learns a mapping between the distribution of high spatiotemporal resolution GCM historic simulations and observations using artificial neural network. We carry out quantitative comparisons against several prior methods to demonstrate the superiority of our approach. We identify GCM biases at grid scale and various spatiotemporal scales.