Observation‐Informed Machine Learning Convective Scheme: A New Training Strategy

Abstract Machine learning (ML)‐based schemes face challenges in online stability and biases because of the accumulated errors during long‐term integrations. Here, we propose a new ML training strategy for the convection process that provides online corrections to tendencies in a more systematic way with significant improvements in stability and bias. The new training strategy employs simulations nudged toward reanalysis dynamical field as targets, while non‐nudging simulations serve as inputs. The results from the low‐resolution hybrid ML‐physics model show that the new ML scheme performs substantially better, as it strengthens heavy stratiform and total precipitation and reduces temperature bias, acting as an effective online bias correction mechanism. In addition, the resulting hybrid model alleviates the long‐standing overestimation of the tropical convective‐to‐total precipitation ratio. Lastly, the hybrid model can run stably for over a year, highlighting the effectiveness and generalizability of the new training strategy.