Tools
Tools
Ravichandran, Naveen Prasad (2025) Hybrid Physics–AI Climate Model with Forcing Awareness for Improved Generalization to Unseen Forcing Scenarios. Masters thesis, Dublin, National College of Ireland.
Preview |
PDF (Master of Science)
Download (4MB) | Preview |
Preview |
PDF (Configuration Manual)
Download (805kB) | Preview |
Abstract
Due to the continuous rise of CO2 in the atmosphere, the task of predicting the climate in the long term with a high degree of reliability becomes more and more challenging. Skilled professionals in the field of atmospheric science constantly utilize machine-learning models to understand the complex climate variations, but these models are very often unable to distinctly predict when the future conditions are different from those during training. One of the main aims of the research is to assess if the CO2 forcing that is made explicit improves the robustness of hybrid climate models. The researchers constructed a unified 1940–2100 dataset that was made up of ERA5 reanalysis, historical NOAA CO2 records, SSP3-7.0 projections, and synthetic trend-preserving extensions. A couple of different architectures were evaluated: one was the standard Temporal Convolutional Network (TCN), and the other was the forcing-aware FiLM-TCN which internally modulates the features using CO2 inputs. The research outcomes confirm that under late-century high-forcing scenarios FiLM-TCN keeps being more accurate and less variable than the others. This highlights the importance of using physical forcing signals in hybrid climate prediction models as a benefit.
Actions (login required)
 |
View Item |