Skip to main content
. 2024 Dec 10;3(12):pgae554. doi: 10.1093/pnasnexus/pgae554

Fig. 5.

Fig. 5.

Scenario with input distribution given in Eq. (7). a) Using the surrogate models trained on related data together with the physics models leads to unbiased multifidelity estimators of the expected burned area that exhibit less variance with increasing computational budget (TPU hours) than using the physics model alone. Including the surrogate model from related data achieves accurate estimates of the expected burned area at already around 10,000 TPU hours, whereas using the physics model alone requires up to almost 50,000 TPU hours to achieve a comparable expected burned area. For a comparison at additional times after ignition; see Fig. S3a. b) Estimates of the RMSEs are in agreement with the previous results and indicate that including the surrogate models trained on related data leads to almost 2× more accurate estimates of the expected burned area compared with using the physics model alone. c) The plot shows the number of samples used from the physics and the surrogate model when the two models are combined by the multifidelity estimator.