Ablation of input variable groups is performed on the surrogate model trained CAM4 data. We measure over skill on a test set as we remove one variable (group) at a time
- Each line corresponds to skill per output variable (note that PTEQ and PTTEND have 26 levels each while PRECC, PRECT are scalars)
- Dashed lines correspond to skill for each output with all input variables
- We observe that OMEGA, Q, T have the greatest impact

Variable names for reference
| Name |
Long Name |
shape |
unit |
| Q |
Specific humidity |
(T, L, 96, 144) |
kg/kg |
| T |
Temperature |
(T, L, 96, 144) |
K |
| U |
Zonal wind |
(T, L, 96, 144) |
m/s |
| V |
Meridional wind |
(T, L, 96, 144) |
m/s |
| OMEGA |
Vertical velocity (pressure) |
(T, L, 96, 144) |
Pa/s |
| PSL |
Sea level pressure |
(T, 96, 144) |
Pa |
| SOLIN |
Solar insolation |
(T, 96, 144) |
W/m2 |
| SHFLX |
Surface sensible heat flux |
(T, 96, 144) |
W/m2 |
| LHFLX |
Surface latent heat flux |
(T, 96, 144) |
W/m2 |
| FSNS |
Net solar flux at surface |
(T, 96, 144) |
W/m2 |
| FLNS |
Net longwave flux at surface |
(T, 96, 144) |
W/m2 |
| FSNT |
Net solar flux at top of model |
(T, 96, 144) |
W/m2 |
| FLNT |
Net longwave flux at top of model |
(T, 96, 144) |
W/m2 |
| Z3 |
Geopotential Height (above sea level) |
(T, L, 96, 144) |
m |
Outputs
| Name |
Long Name |
shape |
unit |
| PRECT |
Total (convective and large-scale) precipitation rate (liq + ice) |
(T, 96, 144) |
m/s |
| PRECC |
Convective precipitation rate (liq + ice) |
(T, 96, 144) |
m/s |
| PTEQ |
Q total physics tendency |
(T, L, 96, 144) |
kg/kg/s |
| PTTEND |
T total physics tendency |
(T, L, 96, 144) |
K/s |