Multi-Year-to-Decadal Temperature Prediction using a Machine Learning Model-Analog Framework

TL;DR

The paper tackles multi-year-to-decadal regional temperature prediction by integrating an analog forecasting framework with a neural network–driven mask of precursors. It builds a CMIP6-based analog library and uses tethering across multiple prior years to constrain matches, producing forecasts for annual $2$-meter temperature with lead times $L$ in $ig[1,10ig]$ years. The main contributions are the learned precursor mask, transfer-learning refinement, and demonstrated performance gains over traditional analog methods and initialized Earth system models across several regions and lead times, with improved mean trends and variability representations. This approach offers a cost-efficient, drift-resilient pathway for probabilistic regional climate projections and provides interpretable insights into which regions act as key precursors for forecast skill.

Abstract

Multi-year-to-decadal climate prediction is a key tool in understanding the range of potential regional and global climate futures. Here, we present a framework that combines machine learning and analog forecasting for predictions on these timescales. A neural network is used to learn a mask, specific to a region and lead time, with global weights based on relative importance as precursors to the evolution of that prediction target. A library of mask-weighted model states, or potential analogs, are then compared to a single mask-weighted observational state. The known future of the best matching potential analogs serve as the prediction for the future of the observational state. We match and predict 2-meter temperature using the Berkeley Earth Surface Temperature dataset for observations, and a set of CMIP6 models as the analog library. We find improved performance over traditional analog methods and initialized decadal predictions.

Figures (18)

• Figure 1: Posthoc selection of best matching analog to observations in three year chunks. This is not usable as a prediction.
• Figure 2: Model learned masks for annual mean 2-meter temperature for five regions: the western United States, the Amazon, northern Europe, the great lakes of Africa, and southern India. Each region (in fact, each grid cell) has a unique mask highlighting global locations that are important markers for the evolution of that region, on that time scale.
• Figure 3: Schematic illustrating the machine learning component of the prediction framework. The learned mask identifies global precursors for predicting a specific variable for a given lead time (L in the figure) for a specific region (western United States in the figure).
• Figure 4: Schematic illustrating the method we use to select analogs. Rather than match only the initialization state to the analog library (blue dashed lines), we enforce that years prior to the initialization also match (yellow solid lines, an example three year match). We call this tethering and it helps avoid selecting by-chance matches that may have very different evolutions.
• Figure 5: Example region-averaged analog predictions and observations for the western United States for a lead time of 5 years. The 50 analog ensemble mean is shown (yellow-black), along with each of the 50 individual analog forecasts (thin, yellow), and the truth from the BEST observational dataset (white-black).