Fig. 6.

Scenario with input distribution given in Eq. (8): Already for low computational budgets of about 10,000 TPU hours, the surrogate model trained on related data together with the physics model lead to a multifidelity estimator of the expected burned area that shows little variance as the budget of TPU hours is increased. This indicates that including the surrogate model from related data provides more accurate estimates than using the physics model alone, because the surrogate model is used via a multifidelity estimator so that it introduces no bias. The estimated RMSEs are in agreement with these results and indicate an almost 3× lower RMSE than the estimator that uses the physics model alone. For a comparison at additional times after ignition, see Fig. S3b.