Hello, A yearly allocation for the LCRC cluster has been requested with the following updated information: Submitter/PI: Robert Jacob Project Name: Climate_Emulator Division: MCS Project title: Building a climate emulator for use in energy policy analysis Associated funding: NSF Other Systems: Science: State-of-the-art general circulation models (GCMs) provide the best estimates of future climate change. While they are valuable tools for assessing impacts of climate change for a given scenario of future greenhouse gas emissions, they cannot be used in iterative economic models since a single multi-century simulation can take weeks to months to run and the iterative models require hundreds of thousands of iterations. For this reason, we aim to develop a statistical emulator that will accurately and efficiently reproduce GCM forecasts of regional climate change for an arbitrary trajectory of CO2 concentrations by using a library of pre-computed GCM output. Library based statistical emulation has not been possible to this date because there is no consistent and comprehensive library of climate model runs corresponding to a range of scenarios of future CO2 concentrations. Computing time on Fusion will enable us to build this unique resource by performing a series of s imulations using the Community Climate System Model (CCSM3). Since library-based statistical emulation is currently not possible, the most common approach to emulating regional climate uses linear pattern scaling, in which changes in regional temperature and precipitation are assumed to be linearly proportional to the change in global mean temperature (GMT). That is, regional climate is effectively assumed to depend solely on GMT, and not on the rate or duration of warming. Our preliminary analysis of CCSM3 output from the Climate Model Intercomparison Project (CMIP3) archive shows that this assumption is not valid for temperature projections over time scales of multiple centuries and breaks down for projections of precipitation even over decadal time scales. The CMIP3 archive, however, is not sufficient for full statistical emulation, since it is both restricted in number of runs and inconsistent in its treatment of regional climate forcing. The carefully designed library we propose here will be used to develop and test a number of alternative approaches for emulating climate model responses, including (i) nearest neighbor interpolation, (ii) a variant of pattern scaling that accounts for the rate of warming, and (iii) principal component analysis (PCA). The intended outcome of this research is a new climate emulator that will aid in the evaluation of energy policies for greenhouse gas mitigation. We anticipate that there will be a wide interest in a user-friendly, computationally tractable emulation tool from the policy, economics, and climate communities. The library itself will also be made public for further use by the community Project description: This request is for the third year of a continuing project on the study and emulation of transient climates – climate after a change in CO2 or other radiative forcing but before the 1000s of years need to reach a new equilibrium. The primary goal of the project has been to build tools to facilitate analysis of climate-related policies. Fusion runs enabled us to pioneer the technique of library-based climate model emulation, in which a library of pre-computed model runs is used to generate a statistical tool that reproduces the output of complex models in a fraction of the time it takes to run the full model. We used Fusion computing time in FY2011 to develop a prototype climate library from runs of CCSM3 (the open-source Community Climate System Model 3 built by the National Center for Atmospheric Research). We demonstrated climate emulation on a subcontinental scale in FY2011 and on at model grid scale in FY2012 (2 papers in preparation) The model runs have also facilitated much synergistic science on the physics of transient climates, including demonstration of the cause of nonlinearities with global mean temperatures for the evolution of both precipitation and global heat uptake. This work directly addresses major outstanding questions in climate science. (1 submitted paper, 2 in preparation). In FY2012 we used much of our Fusion allocation to do several very long (multi-1000-year) climate runs with different forcing scenarios (different sizes or CO2 increases and one long run with solar forcing). Those runs were not discussed in our FY2012 request but proved necessary for the science that emerged from the project, and the timeliness and urgency of this work prompted a reordering of priorities, as these issues are urgent concerns in climate science at present and Fusion computing resources gave us the possibility of answering them. In FY2013 we would like to return to the project we discussed in our FY2012 request, to extend our emulation and study of climate model behavior to cases where climate model parameters are themselves varied. We will continue to work with a relatively low spatial resolution climate model version (CCSM3 at T31, i.e. 3.5ox3.5o resolution) to permit testing more cases and performing longer runs. The dominant computational requirements in CCSM3 are its atmosphere and ocean components. The atmosphere component is the Community Atmosphere Model, which currently supports three different numerical cores for solving the basic equations: a spectral transform method, a Semi-Lagrangian scheme, and a finite-volume scheme. The ocean component is a finite-difference model called the Parallel Ocean Program and was developed at Los Alamos. CCSM3 requires Fortran90 and C compilers and the MPI and NetCDF libraries to run. The study we will return to is the exploration of the importance of parametric uncertainty on climate model projections (especially in parameters relating to cloud coverage and water content, which have been shown to be key drivers of uncertainty in climate projections). The only analogous previous work is the “climateprediction.net” project which used community-donated computing time over many years to do a similar experiment on a higher-resolution version of a British climate model. Lowering the resolution to T31 allows us to be competitive in terms of physical insights with more modest computing demands. In FY2013 we will perform T31 CCSM3 simulations with three values for each of three cloud parameters (nine combinations). For each combination we will spin-up the climate using a 500 year control run and a 140 historical forcing run (1870-2010). Based on results from our FY2011 simulations, we will require three scenarios of future CO2 concentrations from 2010-2160, with three realizations of each scenario, in order to “train” our statistical model. To test the emulator we will require another two scenarios, but will consider only one realization of each. The total requirements for this experiment are 9 * (500 + 140 + 3*150*3 + 2*150) = 21,000 model years. The T31 resolution version of CCSM3 requires 38 core hours per model year, which gives a total of 21,000*38 ~= 800,000 core hours. In FY2012 we requested and received a total allocation of 800K hours, received in two increments, and used all of it. We also received an emergency allocation of 300K hours on August 2; this was a response to the loss of 20% of our climate library in a PADS failure and allowed us to recalculate the runs that were lost. We have approximately 100K left from that allocation and will use all of it by the end of the fiscal year. We have been storing data on the Argonne/UChicago PADS system, which is currently both near capacity and near shutdown. We are therefore in the process of moving to UChicago’s Reseach Computing Center (RCC), which will give us continuing sufficient disk space to store the output offsite. Project URL: Current FY Hours Used: undetermined amount New FY Requested allocation: 800 Q1: 200 Q2: 200 Q3: 200 Q4: 200 Justification: Our allocations so far have allowed us to perform thousands of years of simulations using CCSM3 at T31. This provides us with detailed statistics on the time required to run this model configuration on Fusion. Using 8 nodes (64 processors) takes an average of 38 core-hours for a simulation of one model year. Several simulations can be run at the same time to scale the project out to the full machine if available. We successfully used over 1 million core-hours in FY12. CCSM3 was the recipient of a great deal of tuning and evaluation by combined DOE and NCAR computational scientists so we do not expect further improvements in the per node performance. Thank You, The LCRC Accounts System