LES model and study The training simulations use a modified System for Atmospheric Modeling (SAM) : a non-hydrostatic, anelastic model with prognostic turbulence kinetic energy (1.5-order closure) for unresolved turbulence.
A bin-emulating, two-moment bulk microphysics scheme predicts water mixing ratios and number concentrations separately for cloud droplets and rain (drizzle). Prognostic aerosol number is coupled to cloud microphysics through activation and depleted by collision-coalescence and surface drizzle; sea-salt emissions depend on surface wind. Aerosol sizes follow a single lognormal mode with a geometric mean diameter of 140 nm and geometric standard deviation of 1.5.
Shortwave and longwave radiation are calculated with RRTMG every 20 s. Surface momentum, heat and moisture fluxes use Monin–Obukhov similarity theory. The model formulation is described in Yamaguchi, Zhang, and Feingold (2026, under review), Section 3.1 .
This explorer uses 100 m horizontal grid spacing along MAGIC trajectories in the northeast Pacific: M5 is a precipitation-driven stratocumulus-to-cumulus transition (30 seeded runs); M7 is an entrainment-driven transition (25 seeded runs).
Brightening and the revised decomposition Brightening is the seeded-minus-control change in reflected shortwave radiation at the top of the atmosphere, in W/m². Positive values mean more sunlight reflected to space, not a predicted temperature change.
Our revised method builds on Chun et al. (2023), Appendix B . We adopted a rain-aware decomposition: Nd + cloud water path (CWP) + cloud fraction + residual = total brightening . Nd describes the droplet-number (Twomey) contribution. 2D LES outputs are used to estimate the changes in cloud properties. Cloudy-column changes are estimated using seeded/control ratios at percentiles 1–99. CWP excludes rainwater; the Nd and CWP calculations hold control rain optical depth fixed. Omitted rain changes, interactions, and approximation errors can remain in the residual. This rain-aware treatment is different from the Chun method in attributing Nd and CWP changes, especially under precipitating conditions. More details in Zhang, et al. (2026, in prep).
Time series and post-seed averages Each meteorological case has a joint time-series emulator using principal components and Gaussian processes, plus a separate Gaussian-process emulator for the post-seed time mean. That mean is the time integral divided by elapsed time, including nighttime, so it remains in W/m². It need not equal an integral of the displayed time-series prediction. The residual is derived from the other predictions to preserve the budget.
Reading the controls and map Rate follows the project convention mg⁻¹ s⁻¹. Total amount is rate × duration in seconds, in mg⁻¹—not emitted mass. Times are hours on the simulation clock. The map emphasizes M5 and M7; gray trajectories are not currently emulated. Moving markers follow the saved LES locations, while green segments show the seeding period.
Droplet concentration is separate Nd concentration (cm⁻³) is predicted by an independent emulator and is not the Nd radiative contribution (W/m²). It describes seeded-run droplet concentration in columns with optical depth greater than 2. Short-lived changes can be less accurate than time means.
Accuracy and limits Validation leaves out one complete LES simulation at a time. The calibrated 95% bands describe emulator prediction errors, not all uncertainty in the LES or decomposition View model performance →