Detecting Modeling Problems Early and Quickly

The Science

Weather and climate models are large computer code structures that solve complex systems of mathematical equations. Researchers developed a new, objective, and computationally efficient method to determine whether simulations performed with these codes on new computers or built with new computer software are essentially similar when they don’t match exactly.

The Impact

Weather and climate models that provide important predictions for society are often developed and maintained by a large team of scientists working collaboratively. These teams require objective and efficient testing methods to assure the codes are producing expected behavior as they are being developed in ever-changing computing environments.

Summary

As computer codes are revised, or software and hardware environment are changed, there may be times when it is no longer possible to obtain numbers identical “digit for digit” to previous results. In these situations it is very important, and non-trivial, to distinguish whether these differences are just “noise” or discrepancies caused by unintended coding errors or computing-environment problems. Existing methods that evaluate these discrepancies using long-term statistics of model results are too computationally expensive to use for daily testing during phases of very active model development. A team of researchers led by scientists at Pacific Northwest National Laboratory developed a new method just as robust as existing methods, but hundreds of times cheaper. The new test identifies when simulations performed in a new model or computing environment are considered “changed beyond noise level” by recognizing when the numerical error calculated against a benchmark is found to be inconsistent with previously verified values. The team showed that the new method was effective when applied in the Community Atmosphere Model, and they expect that the underlying concept is generally applicable to atmosphere and geophysical models.

The design of the new test method was inspired by a previous study that evaluated numerical errors related to time evolution in the Community Atmosphere Model.

Principal Investigator(s)

Hui Wan
Pacific Northwest National Laboratory
[email protected]

Phil Rasch
Pacific Northwest National Laboratory
[email protected]

Funding

This research was supported as part of the Accelerated Climate Modeling for Energy (ACME) program, funded by the US Department of Energy, Office of Science, Office of Biological and Environmental Research (BER). High-performance computing resources were provided by the Oak Ridge Leadership Computing Facility at the Oak Ridge National Laboratory, supported by the Office of Science, and the National Center for Atmospheric Research (NCAR) Computational and Information Systems Laboratory.

References

Wan, H., K. Zhang, P. J. Rasch, B. Singh, X. Chen, and J. Edward. “A new and inexpensive non-bit-for-bit solution reproducibility test based on time step convergence (TSC1.0).” Geoscientific Model Development, 10, 537-552 (2017). [DOI: 10.5194/gmd-10-537-2017]