Bridging physics and machine learning: AI-enhanced WRF ensemble approach to extreme rainfall prediction

Extreme weather remains poorly predicted globally due to the rarity, rapid intensification, and inadequate representation in numerical models, thereby limiting forecast skill. With the frequency of extremes rising, reliable forecasts with sufficient lead-time are essential for disaster mitigation. Here, we evaluate rainfall forecasts by integrating AI with Weather Research and Forecasting (WRF) model ensembles during the catastrophic Kerala flood of 14-16 August 2018. Optimized WRF parameterization enhances baseline skill, while En3DVar data assimilation further improves spatial correlation against IMD daily observations from 0.09 to 0.30 (Domain 2, Delta = +0.21) and from 0.10 to 0.69 (Domain 1, Delta = +0.59) for the peak flood day (16 August). We introduce a hybrid WRFDA-Long Short-Term Memory (LSTM) framework that integrates En3DVar-corrected WRF outputs with deep learning post-processing for regional extreme rainfall forecasting, an approach that, to our knowledge, has not been previously demonstrated for convection-permitting ensemble rainfall prediction over complex orographic terrain, which achieves final spatial correlations of 0.69 (Domain 1) and 0.66 (Domain 2) against IMD daily observations, representing total improvements of Delta = +0.59 and Delta = +0.57 respectively over the WRF standalone baseline, with additional LSTM contributions of Delta = +0.00 (Domain 1) and Delta = +0.36 (Domain 2) over WRFDA alone. The LSTM further reduces mean absolute error to 37.90 mm day(-)& sup1; (Domain 1) and 42.73 mm day(-)& sup1; (Domain 2) against IMD, compared to 63.85 and 66.81 mm day(-)& sup1; for the WRF standalone, reductions of 41% and 36% respectively. Nash-Sutcliffe Efficiency improves from -0.88 to 0.44 (Domain 1) and from -1.01 to 0.40 (Domain 2), confirming forecast skill beyond a simple mean reference. Both spatial correlations are statistically significant at the 5% level (p < 0.05; p = 2.52 & times;10(-)(5) and 3.85 & times;10(-)(5) respectively) after accounting for spatial autocorrelation. This fusion of physics-based modelling and deep learning significantly advances extreme rainfall forecasting, offering operationally feasible applicability in tropical regions experiencing rapid climate shifts.