Applying deep learning to a chemistry-climate model for improved ozone prediction

Abstract. Chemistry-climate models have developed significantly over the decades, yet they still exhibit substantial systematic biases in simulating atmospheric composition due to gaps in our understanding of underlying processes. Building on deep learning's success in different domains, we explore its application to correct surface ozone biases in the state-of-the-art chemistry-climate model UKESM1. Six statistical models have been developed, and the model Transformer outperforms others due to its advanced architecture. A simple weighted ensemble approach is further proved to enhance performance by 14 % over the best single model Transformer, reducing RMSE to 0.69 ppb. Applied to future scenarios (SSP3-7.0 and SSP3-7.0-lowNTCF), the UKESM1 shows a larger overestimation of ozone changes by up to 25 ppb compared to present-day conditions. Despite biases, UKESM1 captures the non-linear ozone sensitivity to precursors, with temperature-sensitive processes identified as a dominant contributor to biases. We highlight that simulations of future surface ozone are likely to become less accurate under a warmer climate. Therefore, the bias correction approaches introduced here have substantial potential to improve the accuracy of ozone impact assessments. These methods are also applicable to other chemistry-climate models, which is critical for informing air quality and climate policy decisions.