Improvement of the prediction of surface ozone concentration over conterminous U.S. by a computationally efficient second‐order Rosenbrock solver in CAM4‐Chem

Abstract

The global chemistry‐climate model (CAM4‐Chem) overestimates the surface ozone concentration over the conterminous U.S. (CONUS). Reasons for this positive bias include emission, meteorology, chemical mechanism, and solver. In this study, we explore the last possibility by examining the sensitivity to the numerical methods for solving the chemistry equations. A second‐order Rosenbrock (ROS‐2) solver is implemented in CAM4‐Chem to examine its influence on the surface ozone concentration and the computational performance of the chemistry program. Results show that under the same time step size (1800 s), statistically significant reduction of positive bias is achieved by the ROS‐2 solver. The improvement is as large as 5.2 ppb in Eastern U.S. during summer season. The ROS‐2 solver is shown to reduce the positive bias in Europe and Asia as well, indicating the lower surface ozone concentration over the CONUS predicted by the ROS‐2 solver is not a trade‐off consequence with increasing the ozone concentration at other global regions. In addition, by refining the time step size to 180 s, the first‐order implicit solver does not provide statistically significant improvement of surface ozone concentration. It reveals that the better prediction from the ROS‐2 solver is not only due to its accuracy but also due to its suitability for stiff chemistry equations. As an added benefit, the computation cost of the ROS‐2 solver is almost half of first‐order implicit solver. The improved computational efficiency of the ROS‐2 solver is due to the reuse of the Jacobian matrix and lower upper (LU) factorization during its multistage calculation.

Computational Benefit of GPU Optimization for the Atmospheric Chemistry Modeling

“…Besides the shared memory, there is also 64‐KB constant memory that resides in the GPU's global DRAM memory and can be broadcast among all the SMXs. For the chemistry box model in this study, there are two integer mapping arrays used to extract the 95 reaction‐active species from the total 103 chemical species and permute it to an appropriate order with fewer fill‐in values (Sun et al, ). In the previous implementation, one copy of these arrays is generated for each thread, which is clearly not necessary since all the values in the mapping arrays are constant.…”

“…The first‐order implicit solver is unconditional stable (Kinnison et al, ), but may suffer from low computational efficiency and low accuracy. Sun et al () implemented a second‐order Rosenbrock (ROS‐2) solver in the global chemistry‐climate model (CAM4‐Chem), replacing the original first‐order implicit solver. The results showed that utilizing the same optimized subroutine structure, the ROS‐2 solver achieved ~2× speed‐up on the CPU over the original first‐order implicit solver.…”

“…The first-order implicit solver is unconditional stable (Kinnison et al, 2007), but may suffer from low computational efficiency and low accuracy. Sun et al (2017) implemented a second-order Rosenbrock (ROS-2) solver in the global chemistry-climate model (CAM4-Chem), replacing the original first-order implicit solver.…”

“…In addition, little information was provided to explore the optimal configurations of the CUDA kernels. Hence, in this work, we will port the ROS‐2 chemical solver of CAM4‐Chem (Sun et al, ) to the GPU to solve the chemistry as a box model and examine a series of optimization strategies. The goal is to investigate whether the chemistry box model can benefit from the GPU and its associated most optimized configuration.…”

Maximizing ozone signals among chemical, meteorological, and climatological variability

“…Like many stateof-the-art chemical tracer models, the CAM-chem exhibits some biases, most notably for our purposes a high bias in simulated surface ozone in the eastern US (e.g., Lamarque et al, 2012;Brown-Steiner et al, 2015;Travis et al, 2016;Barnes et al, 2016). Recent efforts have been successful in partially reducing these biases (e.g., Sun et al, 2017).…”