Estimating the Rate of Change of Stratospheric Ozone using Deep Neural Networks

Due to the intensive ozone research in recent decades, the processes that influence stratospheric ozone are well understood . The chemistry and transport model ATLAS was developed to simulate the chemistry and transport of stratospheric ozone globally. The chemical rate of change of ozone is calculated at each model point and time step of the model by solving a system of differential equations that requires 55 input parameters (chemical species, temperatures, ... ). But the computational effort to solve this complex system of differential equations is very high, and with respect to the overall limited computation time, this prevents the inclusion of ozone chemistry into ESMs .

This project proposes a data-driven machine learning approach to predict the rate of change of stratospheric ozone. To derive a data set from modelled data, ATLAS was run for several short model runs. The rate of change of ozone and 55 parameters were stored at each model point and time step. By observing the co-variances of the high-dimensional feature-space, a large data set with reduced dimensionality has been created. A supervised learning algorithm used this data set of input and output pairs to train a deep feed-forward neural network (NN). This involved the identification and optimisation of several hyperparameters and to find a well-functioning combination of depth (number of layers) and width (number of neurons per layer). In this way, the NN model capacity is optimised with respect to the data itself.

To evaluate this approach, the results were compared with another data-driven approach called SWIFT. The SWIFT model employs a repro-modelling approach that uses polynomials to approximate the rate of change of ozone .

The resulting NN model is not only capable of learning the context of an eleven-dimensional hyperplane, but also improves the RMSE by about one order of magnitude compared to SWIFT's previous polynomial approach. In addition, the deviations of the predictions at the boundaries (altitude and latitude) are significantly lower, which is a challenge for the polynomial approach.