Understanding climate and its changes requires the use of complex computer simulations built on dozens of underlying models. The resulting behemoth may predict global climate, but the time and cost of running the models on advanced computers makes it unwieldy to use, especially at fine spatial and temporal resolutions. A group of researchers from Pacific Northwest National Laboratory (PNNL) created and then embedded a physics-informed deep neural network (DNN)—a sophisticated mathematical model that can learn as it processes data—in a detailed regional model of the Amazon rainforest to demonstrate a new way to speed output and resolve compounding errors.
Previous efforts to use DNNs to develop surrogate models for atmospheric phenomena resulted in tiny miscalculations, which can propagate through larger simulations over time until the results don’t represent the realities of nature. Now, a team of PNNL scientists developed a DNN that emulates the complex, multiphase chemistry of a key type of secondary organic aerosol, which can impact climate. Using this DNN, the team showed that they could cut the calculation time for the larger regional model in half, and that it took only 7 hours of data to train the network to generalize results without accumulating errors. This work provides a clear proof of concept for successfully implementing machine learning to replace computationally expensive submodules in climate models, thus speeding up complex physics and chemistry calculations in three-dimensional chemical atmospheric transport and climate models.
Using knowledge of multiphase chemistry of secondary organic aerosols, a team of researchers from PNNL developed and trained a physics-informed DNN to emulate a particular type of aerosol over the Amazon. They used the Tahoma scientific computer at EMSL, the Environmental Molecular Sciences Laboratory, a Department of Energy (DOE) Office of Science user facility, along with other computational resources, to train the DNN on 7 hours of simulations from the Weather Research and Forecasting Model coupled to chemistry (WRF-Chem). Even with such a limited training, the DNN predictions generalized well over several days of WRF-Chem simulations in both the dry and wet seasons of the Amazon. It could successfully simulate the complex composition, spatial and temporal variations, and chemistry of these secondary organic aerosols. Embedding the DNN into WRF-Chem reduces the computational expense of WRF-Chem by a factor of two. The approach shows substantial promise for application to computationally expensive chemistry solvers in climate models.
Manish Shrivastava, Pacific Northwest National Laboratory, firstname.lastname@example.org
The DOE Office of Science, Biological and Environmental Research program supported this research through the Early Career Research Program. The computational resources for the simulations were provided by Pacific Northwest National Laboratory’s Institutional Computing and EMSL, the Environmental Molecular Sciences Laboratory, a DOE Office of Science User Facility. This research also used the computational resources of the Argonne Leadership Computing Facility, which is another DOE Office of Science User Facility.
H. Sharma, M. Shrivastava, and B. Singh, “Physics informed deep neural network as aerosol emulator embedded in a regional chemical transport model.” npj Climate and Atmospheric Sciences 6, 28 (2023). [DOI: 10.1038/s41612-023-00353-y]