Tuning a weather simulation with reinforcement learning.

The WRF Solar Brookhaven model is a real atmospheric physics simulation. It models how clouds affect the sunlight that reaches the ground. The case study is a single day in May 2009, when stratocumulus clouds were observed over the ARM Southern Great Plains site in Oklahoma.

Two parameters control droplet behavior in those clouds: vdis and beta_con. Tuning them by hand or by grid search is impractical because each simulation costs about three hours on six MPI cores in this setup, and the original case study runs about fourteen hours wall clock.

The surrogate idea

Run a bootstrap batch of twenty real WRF evaluations using Latin Hypercube Sampling. Use those points to train a probabilistic MLP that maps parameter pairs to a mean squared error against the ARM SGP observations. The surrogate predicts a mean and a standard deviation, and that standard deviation gets baked into the reward as an upper confidence bonus.

Then SAC trains on the surrogate. Five hundred inner gradient steps. Three hundred greedy rollouts pick the next candidate. One real WRF simulation runs at that candidate. The real result joins the training set. The surrogate is retrained. Repeat.

What worked, and what didn’t

Eighty six WRF runs across the full campaign. Run 1 collapsed: the entropy temperature ran away because the alpha learning rate was too aggressive and the inner loop was too long. The fix was alpha learning rate at 3e-5 and 500 inner steps. Run 3 with a UCB weight at 0.10 pushed the agent into a degenerate low vdis corner for eight of fifteen iterations. Run 4 dropped the UCB weight to 0.02 and stabilized but did not improve on Run 3’s best.

The best run landed at vdis = 0.2108, beta_con = 2.179e22, RMSE 65.05 watts per square meter against the observed irradiance.

This is the DS 493 Capstone with Husain Attarwala, Paridhi Bhardwaj, and Irfan Rathore, advised by Professor Chase Wu. The work is being extended from a conference paper into a journal submission with the team at the NJIT Center for Big Data.