Reconstructing Historical Temperature Fields using Diffusion-based Generative Machine Learning

This study presents advancements in leveraging state-of-the-art computer vision and machine learning techniques to fill historical gaps in temperature field records. Reconstructing historical data is crucial to obtain a complete and precise understanding of past climate scenarios and temperature patterns. Simultaneously, it facilitates more accurate comparisons with present-day climatic conditions, thereby aiding in contextualising the significance of recent years’ record-breaking temperature records. Building upon prior work, our approach aims to enhance the understanding of climate dynamics by training a diffusion-based generative machine learning model that learns the underlying temperature distributions from climate model data. We train a diffusion model on complete (near-surface air) temperature records and condition the generative model on masked fields to reconstruct the missing values (historical gaps). We compare the performance of our methods against statistical and machine learning baselines. We further discuss extensions of our methods to account for temporal correlations and seasonal variability, and aim to include several interpretation methods to validate our results in a more physics-grounded approach.