Numerical weather simulations are an important tool for researching atmospheric processes and understanding the development of complex weather phenomena, such as thunderstorms. High-resolution numerical models allow for a detailed description of the three-dimensional structure of convective systems, but due to their high computational demands, they are often time-consuming. As a result, their use is limited in applications that require fast calculations or real-time simulations. The aim of this thesis was to investigate the possibility of accelerating thunderstorm simulations using machine learning methods while maintaining as much similarity as possible with the results of a numerical atmospheric model.
The Cloud Model 1 (CM1) numerical model was used to generate training data, and it was employed to perform idealized, high-resolution simulations of the development of individual thunderstorms. Meteorological variables describing the three-dimensional structure of the storm, convective development, hydrometeor distribution, and wind field were selected from the CM1 model output. A three-dimensional convolutional neural network with a U-Net architecture was used to predict the further development of the simulation; this network predicts future atmospheric states based on an initial sequence of time steps. Particular attention was paid to the preparation of training data, the division of storm development into individual phases, and the autoregressive generation of sequential time steps. The model’s performance was evaluated through a visual comparison of simulations and using evaluation metrics such as mean squared error (MSE), mean absolute error (MAE), Pearson’s correlation coefficient and the structural similarity index (SSIM).
The results showed that machine learning can significantly speed up the simulation of storm development. The entire simulation, which took approximately 100 minutes to run on the CM1 numerical model, was generated using the developed method in about 45 seconds, representing a speedup of approximately 133 times compared to CM1 and 160 times compared to real time. When compared to CM1 simulations, the model accurately replicated the development of simulations with stable initial atmospheric conditions as well as simulations with unstable initial atmospheric conditions that were similar to the training data. The model successfully preserved the storm’s main structure, the development of convection, and the relationships between individual meteorological variables. Notably, there was a reduced amount of spatial detail, as the model slightly simplified the structures during forecasting. Tests with different initial conditions revealed that the model has limited generalization ability. Changes in terrain and shifts in the initial onset of convection caused minor differences, while changes in the direction of the horizontal wind were not correctly accounted for. In this case, the model followed patterns learned from the training data rather than adapting the storm’s development to the new conditions.
The results confirm that machine learning methods can effectively accelerate high-resolution storm simulations and enable significantly faster generation of future atmospheric conditions. The developed solution serves as a foundation for the further development of such models, in which improved architecture, a larger number of training simulations, and a broader range of initial conditions could lead to greater accuracy and better generalizability to various atmospheric conditions.
|
