Unsupervised Classification of Ozone Profiles from UKESM1

Unsupervised classification of ozone profiles from UKESM1 Unsupervised classification of ozone profiles from UKESM1 Fouzia Fahrin1, Daniel C. Jones2, Yan Wu1, and Alexander Archibald3 1Department of Mathematical Sciences, Georgia Southern University, US 2British Antarctic Survey, NERC, UKRI, UK 3Department of Chemistry, University of Cambridge, UK

Objective Objective The vertical ozone structure in the atmosphere is connected to pollutants and radiative forcing. This distribution features complex spatial and temporal variability, set by balances between chemical reactions, climate variability, and atmospheric transport. We adopted Gaussian Mixture Model (GMM) to study the vertical ozone structure. GMM is an automated, robust, and standardized unsupervised classification technique.

Methods Methods The process of classifying profiles has been done using the following steps: Discard the profiles with NaN values and create a clean data set. Center and standardize data at each vertical level using the mean and standard deviation. Reduce the number of dimensions with Principal Component Analysis (PCA). With 10 principal components, PCA captures 99% of the variance in the vertical structure, which appropriately reduces the number of dimensions from 19 to 10. Set a number of classes K (using the guidance from Bayesian information criterion). Train a classification model using the reduced dimension profiles.

Data Collection Data Collection We use annual mean ozone profiles from three different experiments from the UK Earth System Model 1 (UKESM1). Historical Annual means covering 2004-2014. SSP126 Annual means covering the years 2090-2100 with strong greenhouse emission reduction. SSP585 Annual means covering the years 2090-21000 with no greenhouse emission reduction.

Ozone is the most abundant in the lower and middle stratosphere and decreases gradually in the upper stratosphere. Figure: Vertical Ozone Structure from UKESM1

Bayesian Bayesian information information criterion criterion (BIC) (BIC) We chose the number of classes K based on our desired level of complexity of the GMM. The BIC metric rewards increased likelihood and penalizes overfitting. The variation in the BIC score comes from sampling 678810 different random samples of 1000 profiles using same random seed in 20 iterations at each K value. We found a range of values of K that are consistent with small BIC scores. Choosing from this range, we selected K=8 for ease of interpretation. Ideally, we want to select the K value that minimizes BIC. Shown: Mean +/- one standard deviation

Gaussian Mixture Model (GMM) identifies different profile types. These profile types are from all three UKESM data sets we used in the study. These classes are sorted by mean latitude. To ensure that we train GMM using a comprehensive dataset, we selected training data from all three UKESM1 experiments. Figure: Ozone concentration statistics of UKESM profiles, separated by class, as a function of pressure.

Some observations Classes at both poles feature a low tropopause (Class 1, 2, 3, 7, and 8). Except for the tropical classes, all the other classes have relatively large standard deviation in lower stratosphere. All the classes have posterior probability more than 90%. The surface ozone is higher in the Northern Hemisphere mid-latitude classes (especially class 6). Arctic classes have higher stratospheric ozone than Antarctic classes.

Without latitude and longitude information, Gaussian mixture model identifies spatially distinct regions. Class 8 does not appear in the historical data. Figure: Map of profiles color-coded with the class they have been attributed to for model historical profiles.

Tropical classes (class 4 and 5) switched place from the Historical label Map. Class 2 does not appear in the SSP126 data. Figure: Map of profiles color-coded with the class they have been attributed to for SSP126 profiles.

The spatial extent varies for each model data set. Figure: Map of profiles color-coded with the class they have been attributed to for SSP585 profiles.

Class Hemisphere mid-latitudes has high surface ozone in both historical and SSP585, indicating the influence of wildfires in that region, which act as a significant precursor. 6 over the Northern surface ozone SSP126 has low surface ozone because ozone precursors emission is strongly reduced in this case. Southern have low surface ozone than Northern Hemisphere classes. Hemisphere classes

The tropical classes expand in area in both SSP126 and SSP585, at the expense of the higher-latitude classes Class Historical SSP126 SSP585 Historical SSP126 SSP585 1 7.50 10.80 7.05 Southern High latitude Classes (Class 1, 2, 3) 29.8% 26.57% 28.43% 2 8.00 1.59 8.96 3 14.30 14.18 12.42 4 19.70 28.64 24.67 Tropical Classes (Class 4, 5) 42.8% 47.8% 46.20% 5 23.11 19.16 21.53 6 13.40 12.12 11.82 Northern High latitude Classes (Class 6, 7, 8) 27.4% 25.64% 25.36% 7 12.31 11.12 13.50 8 1.70 2.40 0.04 Table: Relative area occupied by classes, in percentages of the total area.

Conclusions In this study, Gaussian Mixture Modeling (GMM), an unsupervised classification method, has been used with UKESM1 model data to identify common structures in atmospheric ozone profiles robustly and objectively. Although no latitude or longitude information was included in the GMM, each class corresponds to a specific, coherent region. We trained the GMM using data from all three model cases for easier comparison. Unsupervised classification uses a limited number of classes and spatial information to reveal robust ozone structures. We hope this study demonstrates the potential of unsupervised machine learning in the study of ozone structure in numerical models.