Filling N2O data gaps during the last 800,000 years via Artificial Intelligence Techniques

Abstract

One of the primary proxies for ancient atmospheric air compositions is the fossil air occluded in polar ice sheets. Ice cores are significant archives for the atmospheric Greenhouse Gas (GHG) concentrations during the last 800 kyr (thousand years). GHG records from polar ice cores have provided valuable information on past climatic, atmospheric, and glaciological changes. For example, nitrous oxide (N2O) is a long-lived GHG and gives us information on nitrogen cycles. However, the time resolution and missing gaps of N2O records limit our understanding of the control mechanisms in the atmosphere. One of the well-known state-of-the-art methods is AI (Artificial Intelligence) techniques, and its main branch is ML (Machine Learning) method. To fill the N2O gaps during the last 800 kyr, we used CO2 and CH4 concentration data from Antarctic ice cores (Vostok and EPICA Dome C ice cores). The ML methods were run in two steps. First, the gap parts were deleted from the time series, and modeling was performed with CO2 and CH4 concentration data with six various ML methods (Linear, Support Vector Machine, Tree, Gaussian Process Regression (GPR), Artificial Neural Network (ANN), and Ensemble). Then, the best model was selected for the second step to reconstruct the N2O in the data gaps. Although other AI methods revealed acceptable results in the first step, the GPR method produced a high-quality simulation with R2= 0.86, RMSE= 7.38 ppb, and MAE= 4.15 ppb. Finally, the simulation for the N2O gaps was performed with the GPR method. Our results confirm that AI techniques have a high potential to produce continuous paleo atmospheric GHG concentrations. Future research includes modeling with other AI and ML methods and applying the AI techniques to other ice core records, such as water isotope ratios (temperature proxy) over various past periods.