CCCma-PPP: A Machine Learning Post-Processing Framework for Seasonal-to-Decadal Climate Predictions

A machine learning pipeline developed by the Canadian Centre for Climate Modelling and Analysis (CCCma) for post-processing seasonal-to-decadal climate predictions, supporting tasks such as bias adjustment, ensemble boosting, and statistical downscaling.

• Original Author/Maintainer: Parsa Gooya (Environment and Climate Change Canada, Canadian Centre for Climate Modelling and Analysis)
• Source Platform: GitHub
• Original Title: CCCma-PPP
• Original Link: https://github.com/ParsaGooya/CCCma-PPP
• Publication Time: 2026

---

Climate prediction is a core tool for understanding the evolution of the Earth system and addressing the challenges of climate change. However, the output data from Global Climate Models (GCMs) often has systematic biases, limited spatial resolution, and high uncertainty in single-model predictions. These issues limit the value of climate predictions in practical applications such as agricultural planning, water resource management, and disaster early warning.

To improve the accuracy and usability of climate predictions, scientists have invested significant research in the field of model output post-processing. Post-processing techniques can effectively reduce biases, improve resolution, and enhance prediction reliability by statistically correcting and optimizing raw model outputs. While traditional post-processing methods have achieved some success, machine learning methods show greater potential when dealing with massive data and complex patterns.

CCCma-PPP is a machine learning pipeline developed by the Canadian Centre for Climate Modelling and Analysis (CCCma), specifically designed for post-processing seasonal-to-decadal predictions generated by the Canadian Earth System Model (CanESM). This project represents cutting-edge practice at the intersection of climate science and machine learning, providing a systematic solution for the refined processing of climate prediction data.

The core design goals of this framework are to support multiple post-processing tasks, including:

• Bias Adjustment: Eliminate systematic biases in climate model outputs
• Ensemble Boosting: Integrate information from multiple models or ensemble members to improve prediction robustness
• Statistical Downscaling: Downscale coarse-resolution climate predictions to finer local scales

The training framework of CCCma-PPP is highly flexible, supporting the mixed use of deterministic and probabilistic models. At the neural network architecture level, the framework is compatible with multiple mainstream architectures, allowing researchers to select the most appropriate model structure based on specific application scenarios.

This multi-architecture support capability is of great significance. Different climate variables (such as temperature, precipitation, sea level pressure) and different prediction time scales (seasonal, interannual, decadal) often exhibit different statistical characteristics, so targeted selection of model architectures is necessary. For example, some variables may be more suitable for using recurrent neural networks to capture temporal dependencies, while spatial downscaling tasks may benefit from the structure of convolutional neural networks or graph neural networks.

The practical application value of this project is reflected in multiple aspects:

For climate researchers, CCCma-PPP provides a standardized post-processing toolchain that can significantly reduce repetitive development work, allowing researchers to focus more energy on the scientific problems themselves. At the same time, the modular design of the framework facilitates the expansion of new post-processing algorithms and promotes methodological innovation.

Meteorological and hydrological operational agencies can use this framework to improve the quality of climate prediction products. For example, the agricultural sector can obtain more accurate seasonal precipitation predictions to optimize planting decisions; water resource management departments can obtain more reliable runoff predictions to improve reservoir operation efficiency.