Physically Constrained Artificial Intelligence General Circulation Model (AIGCM) for Climate Prediction

Introduction

The rise of artificial intelligence (AI) and machine learning (ML) in climate science has led to significant advancements in predictive accuracy and computational efficiency. However, pure ML-based models often suffer from biases that can misrepresent long-term climate trends. To address this issue, the concept of a physically constrained ML model can play an important role.

The Importance of a Good Mean State

When working with climate modeling, it is essential to have a reliable grasp of the average climate state. If the mean state is not accurate, whether using machine learning or dynamical models, systematic errors are more likely to occur. A dependable understanding of the mean state is crucial to comprehend and predict variations, trends, and anomalies.

The Pitfalls of Unconstrained ML Models

While ML models excel in capturing complex, nonlinear relationships in data, they can often suffer from what's known as mean state model bias. This bias can come from numerous sources, including data limitations and underlying assumptions. While ML models may be great at short-term anomaly prediction/forecast, their predictive skills often deteriorate over longer time scales due to this inherent physical bias.

Physically Constrained ML Models to the Rescue

To overcome the limitations of traditional ML models in climate science, physically constrained ML models can play a key role. These models integrate physical equations and constraints from dynamical systems, helping to guide the learning process. The result is a hybrid model that maintains the efficiency and adaptability of ML models while benefiting from the physically accurate representation provided by dynamical models.

Artificial Intelligence General Circulation Model or AIGCM

I have developed the Artificial Intelligence General Circulation Model (AIGCM), which is a physically constrained ML model that is still in a very early stage. To illustrate the potential of physically constrained ML models, consider a free-running climate model simulation from AIGCM. In this case, a Neural Net model was trained on reanalysis data, and the free-running forecast was done using only the initial conditions as a starting point. The model was designed to predict subsequent time steps based on its own previous predictions, running autonomously without the need for observed data input.

Results and Observations

What was observed was a significantly reduced bias compared to just the ML model. The long-term projections were more consistent with established climate trends, thereby improving the model's reliability for climate change forecasting (Fig. 1). Our simulation was designed to predict the same year over and over again, so there shouldn't be any annual mean trends other than interannual variability. However, due to the model's inherent bias, the ML model starts warming very quickly without any climate change signal. Whereas, the physically constrained ML model simulates the seasonal cycle correctly without the annual mean trend of the temperature.

Comments

Physically constrained ML models offer a promising pathway for improving the accuracy and reliability of climate predictions. As computational capabilities continue to grow, there is an increasing opportunity to further refine these models for more accurate and reliable long-term projections.