A machine learning based deep convective trigger for climate models

The present study focuses on addressing the issue of too frequent triggers of deep convection in climate models, which are primarily based on physics-based classical trigger functions such as convective available potential energy (CAPE) or cloud work function (CWF). To overcome this problem, the study proposes using machine learning (ML) based deep convective triggers as an alternative. The deep convective trigger is formulated as a binary classification problem, where the goal is to predict whether deep convection will occur or not. Two elementary classification algorithms, namely support vector machines and neural networks, are adopted in this study. Additionally, a novel method is proposed to rank the importance of input variables for the classification problem, which may aid in understanding the underlying mechanisms and factors influencing deep convection. The accuracy of the ML-based methods is compared with the widely used convective available potential energy (CAPE)-based and dynamic generation of CAPE (dCAPE) trigger function found in many convective parameterization schemes. Results demonstrate that the elementary machine learning-based algorithms can outperform the classical CAPE-based triggers, indicating the potential effectiveness of ML-based approaches in dealing with this issue. Furthermore, a method based on the Mahalanobis distance is presented for binary classification, which is easy to interpret and implement. The Mahalanobis distance-based approach shows accuracy comparable to other ML-based methods, suggesting its viability as an alternative method for deep convective triggers. By correcting for deep convective triggers using ML-based approaches, the study proposes a possible solution to improve the probability density of rain in the climate model. This improvement may help overcome the issue of excessive drizzle often observed in many climate models.