Fig. 2: Objective function (root-mean-square error, RMSE) and computational time for the testing period vs. number of forward runs per gridcell. | Nature Communications

Fig. 2: Objective function (root-mean-square error, RMSE) and computational time for the testing period vs. number of forward runs per gridcell.

From: From calibration to parameter learning: Harnessing the scaling effects of big data in geoscientific modeling

Fig. 2

Here, a forward run for dPL means running through the same number of days as the specified training period, e.g., 30 random instances of 240-day-long simulations in a minibatch would be counted as ~20 runs for a 2-year training period. Dashed lines are for SCE-UA and solid lines are for dPL. s16, s8, and s4 denote models trained with sampling densities of 1/162, 1/82, and 1/42, respectively, where 1/162 represents sampling one gridcell from each 16 × 16 patch. The dotted horizontal line represents the RMSE threshold of 0.05, below which a model is considered functional. a RMSE for the models trained with 1 year’s worth of data. b Same as a but for models trained with 2 years’ worth of data. c RMSE vs computational time per gridcell. Both methods use the same surrogate model running on one GPU, and 1 year’s worth of training data. d Same as c but for models trained with 2 years’ worth of data.

Back to article page