Improving Indian summer monsoon rainfall prediction using deep learning up to two years in advance

Abstract Long‐lead seasonal forecasts (>12 months) of the Indian summer monsoon rainfall (ISMR) are crucial for adaptive planning and damage minimization against climate change‐induced increasing threats of higher frequency of hydrological disasters in the coming decades. However, the growth of initial errors and drift of forecast climatology with lead month drive the seasonal forecast skill of ISMR by Atmosphere–Ocean General Circulation Models (AOGCMs) to decrease with lead time, making them useless beyond an approx. six‐month lead. Hope of overcoming the challenge is rekindled from recent advances in the application of deep‐learning models to weather and climate prediction that extend the skill of weather prediction beyond the limit of the best numerical weather prediction (NWP) models. Simultaneous to this advance, we have established the physical basis of high‐potential predictability of seasonal forecast of ISMR up to 24‐month leads. Here, we develop a physics‐guided deep‐learning (PGDL) model‐based ‘long‐lead forecast system’ for ISMR trained on the relationship between ISMR and the depth of the 20 °C isotherm in the tropics from a large ensemble of AOGCM historical simulations and past observations to overcome the challenge of poor forecast skill of ISMR at long leads. In contrast to the initialized physical AOGCMs, our model makes increasingly skillful seasonal forecasts up to 24‐month leads in accordance with potential predictability while demonstrating superior skill in predicting extreme excess/deficient ISMR between 1980 and 2023 at 18‐ and five‐month leads. Operational feasibility of the model is demonstrated by making an experimental 18‐month lead forecast of ISMR for 2024. Our findings establish a physical basis and methodology for long‐lead seasonal prediction of ISMR and a building block for three‐dimensional tropical seasonal predictions.