Control simulation experiments using a differentiable global climate model: A case study of Typhoon Jebi (2018)

Control simulation experiments (CSEs) aim to steer model predictions toward a desired target by intervening in the model inputs. This concept is mathematically analogous to data assimilation, in which model states are adjusted to align with observations. In recent machine-learning models, automatic differentiation is used to compute gradients of a loss function and to optimize model parameters to minimize the difference between predictions and targets (e.g., reanalysis data). By optimizing input variables (initial conditions) instead of model parameters, CSEs can be formulated for differentiable models. In this study, we conduct CSEs for Typhoon Jebi (2018) using NeuralGCM, a fully differentiable global climate model. Tropospheric temperature perturbations are optimized to minimize the mismatch of 500-hPa geopotential height (Z500) over 130°E–160°E and 10°N–40°N. The results demonstrate that the predicted fields converge toward the target, indicating that CSEs can be successfully implemented and executed in a differentiable GCM framework.