Deep Learning Model Pretrained on Simulations Predicts Monthly Global Temperature Fields

A deep learning model pretrained on extensive climate simulations can predict monthly global near-surface temperature fields, presenting a computationally efficient alternative for scientific applications that require repeated numerical integrations. Reporting in the journal Theoretical and Applied Climatology, researchers Xuyuan Guan, Jingyun Zheng, and Xuezhen Zhang addressed a core obstacle in applying advanced machine learning to climate prediction: historical observational reanalysis datasets are relatively short compared to the massive data volumes necessary to properly train high-capacity deep neural networks. To resolve this observational data bottleneck, the authors constructed a simulation-pretrained deep learning system named the Forecasting Climate Model, designed to leverage long-term climate model simulations before transferring that learned physical knowledge to real-world observational contexts.

The Forecasting Climate Model is structured to generate monthly global forecasts using eight consecutive months of historical climate inputs. Specifically, the architecture ingests eight months of global near-surface temperature fields, atmospheric carbon dioxide concentrations, and volcanic aerosol optical depth measured at 550 nanometers to forecast temperature conditions for the subsequent month. To build the foundational representations of the model, the research team trained the network on extensive Coupled Model Intercomparison Project Phase 6 past1000 and historical simulations. For evaluation, the researchers tested the architecture on simulations from the MIROC-ES2L model, which was kept completely separate and excluded from all model development procedures, while establishing separate ERA5 reanalysis experiments to examine how well simulation pretraining transfers to reanalysis data.

Evaluations conducted on the independent MIROC-ES2L test set demonstrated that the Forecasting Climate Model achieved high predictive skill for one-month forecast horizons. On this unseen climate model dataset, the network recorded a one-month area-weighted root mean square error of 1.12 Kelvin across the globe. This predictive performance marked a substantial improvement over standard baseline approaches evaluated under identical test conditions. For example, an anomaly persistence benchmark recorded an area-weighted root mean square error of 1.49 Kelvin, whereas raw persistence registered an error of 3.22 Kelvin. Furthermore, predictions calculated from the training climatology yielded an area-weighted root mean square error of 3.01 Kelvin, illustrating that the simulation-pretrained architecture learned time-evolving dynamical patterns rather than relying on static seasonal averages.

Despite the deep learning model’s advantages over persistence baselines, a forcing-conditioned lag-8 Linear Inverse Model proved to be a highly competitive statistical alternative. In one-step evaluations, the Linear Inverse Model achieved a slightly lower area-weighted root mean square error of 1.077 Kelvin compared to 1.124 Kelvin for the Forecasting Climate Model. Additionally, the linear inverse formulation achieved superior spatial pattern accuracy, registering a spatial anomaly correlation coefficient of 0.326 compared to 0.210 for the deep learning model. These comparative results indicate that established linear inverse methods can match or slightly exceed the single-step statistical accuracy and spatial correlation of deep learning systems when conditioned on identical external forcings.

The researchers further evaluated the long-term robustness of both modeling frameworks by conducting extended 360-month recursive integrations driven by prescribed atmospheric carbon dioxide and volcanic aerosol optical depth forcings. Throughout these 30-year recursive simulations, both the Forecasting Climate Model and the Linear Inverse Model maintained numerical stability and substantially outperformed climatological forecasts. Neither model exhibited persistent superiority across all lead times during these extended multi-decade runs. However, structural diagnostics revealed important physical constraints in the deep learning forecasts: the Forecasting Climate Model systematically underestimated temperature anomaly variance and concentrated an excessive degree of variability into its leading empirical orthogonal function modes.

To understand the internal mechanisms of the Forecasting Climate Model and its practical utility for observational data, the authors carried out ablation tests, perturbation experiments, and transfer learning evaluations. Carbon dioxide ablations revealed that carbon dioxide information and the model’s explicit projector pathway provided a modest but measurable contribution to forecast skill, while idealized aerosol optical depth perturbations at 550 nanometers generated systematic, forcing-dependent responses. In transfer experiments applying the model to ERA5 reanalysis, progressive unfreezing achieved the best performance under matched optimization settings, recording a full-field root mean square error of 1.232 Kelvin, compared with 1.353 Kelvin for full unfreezing and 2.041 Kelvin for training on ERA5 data alone. The findings support simulation pretraining as an effective strategy for monthly temperature prediction while identifying variance retention and higher-order structural variability as vital limitations of deterministic deep-learning forecasts.

Subject of Research: Deep learning forecasting of monthly global temperature fields

Article Title: A simulation-pretrained deep learning model for monthly global temperature-field prediction

Article References: Guan, X., Zheng, J., & Zhang, X. (2026). A simulation-pretrained deep learning model for monthly global temperature-field prediction. Theoretical and Applied Climatology, 157(12), Article 789. https://doi.org/10.1007/s00704-026-06710-0

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06710-0

Keywords: climate modeling, deep learning, temperature forecasting, climate simulations, reanalysis data, machine learning, atmospheric science, Linear Inverse Model, simulation-pretrained, deep, learning, model