<p>Climate variability on a monthly to annual scale is the foundation for high-impact climate anomalies and extreme weather that pose a threat to a range of sectors in our society. With anthropogenic global warming, climate variability is projected to change non-linearly. The consequences and impacts of those changes already stretch around the globe. Unfortunately, current generation coupled-climate models still lack skill in representing variability in crucial parts of the climate system. Gaining more insights about a realistic range of climate variability is therefore of utmost importance. Recently, rapid progress was made in the implementation of artificial intelligence tools for climate science. Especially deep learning tools show promise in extracting features of interest out of gridded data, forecasting time series and representing physical systems.</p><p>Here we present a basic neural network approach that reconstructs global fields from sparse local data. Rather meant as a proof-of-concept than the best possible climate reconstruction, we stay extremely conservative in both the spatial and temporal availability of data, and as such operate on very small sample sizes for a deep learning approach. Although our goal is to reconstruct temperature anomalies, our approach is flexible concerning the variable reconstructed, local data coordinates and input data type. In support of the United Nations Sustainability Goals, training and generating of our final, 4800 month temperature anomalies reconstruction cost around 30 minutes on a middle-class, GPU-supported laptop.</p><p>We focus here on a grid reconstruction problem, a typical issue in the (paleo-) climate community. By using a very conservative approach with small sample sizes (especially for a deep learning approach), we could produce a realistic, performant global temperature anomaly reconstruction. Compared to previous gridded climate reconstruction approaches, we use very scarce point data to reconstruct four orders of magnitude more data.</p>