Child pages
  • Abstract: Simulating global climate with turbulence-permitting cloud superparameterization to train machine learning emulators and advance understanding of aerosol-cloud feedbacks, Y2
Skip to end of metadata
Go to start of metadata

Earth's global temperature responds to changes in the balance between incoming solar radiation and outgoing longwave radiation. The largest uncertainty in quantifying the radiative impacts of human emissions on the present and historical climate is caused by the interaction of aerosols with clouds. From a computational modeling standpoint, the main challenge is the multi-scale nature of the problem. On the one hand, conventional global climate models rely heavily on suboptimal parameterization of cloud-scale processes. On the other hand, global simulations that resolve cloud-scale processes are still prohibitively expensive.

In this XRAC renewal proposal, we address this problem by combining the strengths of a multi-scale model and machine learning techniques to better advance our understanding of aerosol-cloud interactions. Building on recent success in our pilot XRAC, our strategy is to run the multi-scale Ultra-Parameterized Community Atmosphere Model (UPCAM) that embeds large-eddy resolution cloud resolving models within larger domain climate model columns and train a neural network with the model output to leverage machine learning (ML) techniques to estimate the impact of aerosols on cloud radiative properties. First, we will catalogue present-day aerosol-cloud relationships using UPCAM simulations. Second, we will compare simulations with present-day and pre-industrial emissions of aerosols to quantify the impact of anthropogenic aerosol emissions on cloud properties. Third, a spectrum of neural network architectures will be trained on the output from the UPCAM simulations, to test the viability of modern machine learning – including emerging physics-constrained generative techniques - for predicting the effect of anthropogenic aerosol emissions on cloud properties. A key innovation in the renewal proposal is the use of new multi-grid and geographic load-balancing techniques to accommodate increasingly faithful benchmark simulations of the aerosol-cloud interaction at ambitious computational scale, with downstream benefits for the ML workflow trained on these data. Several other innovations are explored in the context of the rapidly moving frontier of machine-learning based emulation of subgrid turbulence, not only for climate modeling but also for physical understanding.

  • No labels