Numerical weather prediction with the Weather Research and Forecasting (WRF) model depends critically on physics parameterization choices. For tropical cyclone forecasting over the Bay of Bengal, default parameter values are not tuned for regional conditions, leading to systematic errors in wind speed and precipitation. Exhaustive sensitivity analysis is feasible but computationally expensive when performed through direct simulation.
This study applies machine learning-based multi-objective optimization to calibrate eight sensitive WRF physics parameters simultaneously. The calibration dataset covers ten Bay of Bengal cyclone events from 2011–2017, targeting errors in 10 m wind speed and precipitation as dual objectives. A surrogate model replaces direct WRF runs during optimization, enabling efficient exploration of the high-dimensional parameter space at a fraction of the computational cost.
The calibrated parameter set improved 10 m wind speed prediction by 17.6% and precipitation by 8.2% relative to WRF defaults. Findings were validated across independent cyclone events and tested for parameter robustness across varying storm conditions, demonstrating that ML-guided calibration can yield physically meaningful and transferable improvements to regional NWP.