Understanding monthly-to-annual climate variability is crucial for adapting to future climate extremes. This paper presents a machine learning-based method using a Recurrent Neural Network (RNN) to reconstruct climate variability. The method utilizes sparse, realistically distributed pseudo-station data and demonstrates realistic temperature patterns and magnitude reproduction. The reconstruction process, including training and generation of over 4800 months of global temperature anomalies, is computationally efficient, taking approximately one hour on a standard laptop. The method is adaptable for various regions, periods, and variables.

Martin Wegmann, Fernando Jaume-Santero