Anticipating climate impacts and risks in present or future climates
requires predicting the statistics of high-impact weather events at
fine-scales. Direct numerical simulations of fine-scale weather are
computationally too expensive for many uses. While regression-based
(deep-learning or statistical) downscaling of low-resolution climate
simulations is several orders of magnitude faster than direct numerical
simulations, it suffers from several limitations. These limitations
include the tendency to regress to the mean, which produces excessively
smooth predictions and underestimates the magnitude of extreme events.
Additionally, they also fail to preserve statistical measures that are
key for climate research. We use a conditional GAN (c-GAN) architecture
to downscale daily precipitation as a Regional Climate Model (RCM)
emulator. The c-GAN generates plausible residuals on top of the
predictable expectation state produced by a regression-based DL
algorithm. The skill of c-GANs is highly sensitive to a hyperparameter
known as the weight of the adversarial loss (λadv), and
the value of λadv required for accurate results varies
with season and performance metric, casting doubt on the robustness of
c-GANs as usually implemented. But, by applying a simple intensity
constraint to the loss function, it is possible to obtain robust
performance results across λadv spanning two orders of
magnitude. C-GANs are considerably more skillful in capturing
climatological statistics including the distribution and spatial
characteristics of extreme events. We expect c-GANs with this
modification to be readily transferable to other problems and time
periods, making them a useful weather generator for representing extreme
event statistics in present and future climates.