Abstract
The next generation of Earth System models promisesunprecedented
predictive power through the application of improvedphysical
representations, data collection, and high-performancecomputing. A key
component to the accuracy, efficiency, and robustnessof the Earth System
simulations is the time integration ofdifferential equations describing
the physical processes. Manyexisting Earth System models are simulated
using low-order,constant-stepsize time-integration methods with no error
control,opening them up to being inaccurate, inefficient, or require
aninfeasible amount of manual tweaking when run over
multipleheterogeneous domains or scales. We have implemented the
variable-stepize, variable-order differentialequation solver SUNDIALS as
the time integrator within the Structurefor Unifying Multiple Modelling
Alternatives (SUMMA) modelframework. The model equations in SUMMA were
modified and augmented toexpress conservation of mass and enthalpy.
Water and energy balanceerrors were tracked and kept below a strict
tolerance. The resultingSUMMA-SUNDIALS software was successfully run in
a fully automatedfashion to simulate hydrological processes on the North
Americancontinent, sub-divided into over 500,000 catchments. We compared
the performance of SUMMA-SUNDIALS with a version (calledSUMMA-BE) that
used the backward Euler method with a fixed stepsize asthe
time-integration method. We find that SUMMA-BE required two ordersof
magnitude more CPU time to produce solutions of comparable accuracyto
SUMMA-SUNDIALS. Solutions obtained with SUMMA-BE in a similar orshorter
amount of CPU time than SUMMA-SUNDIALS often contained
largediscrepancies. We conclude that sufficient accuracy, efficiency,
and robustness ofnext-generation Earth System model simulations can
realistically onlybe obtained through the use of adaptive solvers.
Furthermore, wesuggest simulations produced with low-order,
constant-stepsizesolvers deserve more scrutiny in terms of their
accuracy.