Continuous gap-filled atmospheric N2O record for the past 800,000 years using machine learning techniques

Abstract

Ice cores are crucial archives of atmospheric greenhouse gas (GHG) concentrations. Despite the importance of nitrous oxide (N2O) as a GHG, existing ice core records contain gaps, particularly during glacial periods, due to the high dust content in ice samples that may cause in situ chemical or biological reactions, increasing N2O concentration. By developing an iterative process that applies machine learning (ML) models to existing data on CO2, CH4, and N2O from Antarctic ice cores, we simulated a continuous time series of atmospheric N2O concentrations for the past 800,000 years (kyr). The continuous N2O record allows us to investigate long-term variability and potential climate feedback that would otherwise remain obscured, as spectral analysis of this record has revealed significant N2O periodicities of ~100, 41, and 23 kyr. While ML-based simulations cannot fully replace real, artifact-free measurements, they provide a valuable complementary approach to interpreting past climate dynamics, especially when empirical data are limited or compromised.