확률론적 물리 모수화 개발 및 물리 모수화에서의 기계학습 활용

While there are many challenges associated with developing physics parameterization, one major issue is developing physics parameterization that can accurately represent (or parameterize) the stochastic nature of subgrid variabilities. A stochastic convection scheme was developed with convective updraft plumes at the surface randomly sampled from the correlated multivariate Gaussian distribution for updraft vertical velocity and thermodynamic scalars. In addition to the stochastic initialization at the near-surface, a stochastic mixing model with a machine learning technique is proposed. A stochastic convection parameterization is tested using a single-column model and a global climate model, and the simulation results indicate that the parameterization improves the simulation of intraseasonal variabilities. We are also testing PDF-based turbulence parameterization using the Lagrangian stochastic modeling approach, which can fully simulate the PDFs of subgrid variabilities. This seminar will also discuss recent advances in machine learning based physics parameterization and their potential applications.