Hello, A new project on the LCRC cluster has been requested. Please forward the information on to the LCRC Allocation sub-committee. Applicant's name: Spencer Hills Applicant's institution: ANL Applicant's division: NST Project Name: NanoDet Project title: Combined Experimental-Computational Nanocluster Structural Determination Associated funding: LDRD Other Systems: Carbon at Center for Nanoscale Materials (620,000 hours) Cori at NERSC (260,000 hours) Science: Determining the atomistic structures of nanoscale systems such as nanoclusters is a fundamental problem in nanoscience. The knowledge of these structures reveals important information about the functionality of these systems, allowing them to be used for applications such as opto-electrics and catalysis. Although there are both experimental and computational methods to determine these nanoscale structures, they both possess limitations which demonstrate the difficulty of nanoscale structural determination. With experimental data, such as pair distribution function (PDF), multiple possible structures may correspond to the same data, but not all such structures are physically plausible. In computational work, theoretical calculations of the system may result in several structures with similar energies, as well as requiring numerous configurations of the system to be tested. Using a modular computational framework currently being developed at the Theory and Modeling group of the Center for Nanoscale Materials, we expect to overcome the limitations of either data type by combining both experimental and computational data. This will allow us to efficiently sample the possible configurations of our test systems and determine the atomistic structures of various systems, from nanoclusters to grain-boundaries and PV interfaces. Project description: We will use the modular global optimization code which is being developed in Python by current members of our group (Spencer Hills, Fatih Sen, Grace Lu, Eric Schwenker, Maria Chan). This code can perform several types of global optimization, including GA for periodic and non-periodic systems. We plan to consider several systems within the project: 1. The structure of transition metal nanoclusters, such as gold nanoclusters, which are useful for catalytic processes. These structures have been studied extensively by our group with plane-wave DFT calculations (VASP) using a genetic algorithm, which found a broad range of structures that exhibit very large structural differences with only small energy differences. These systems represent a perfect test case for the global optimization code that we are developing, since they have diverse geometries with minimal energy differences, which allows us to test the ability of the code to distinguish between similar energy structures. We will perform global optimization of the structures using combined experimental-computational optimization and only computational optimization to compare the efficiency of each method. We will use simulated pair distribution function (PDF) data from DiffPy for the experimental data and DFT calculations using VASP for computational data. For each glo bal optimization, in this case genetic algorithm, we need to perform DFT calculations for 1000 structures. Assuming 15 minutes for each structure to relax on 64 cores, 16,000 hours are required per optimization. In order to produce meaningful comparison between the two sets of optimizations, we will run 20 optimizations (10 combined experimental-computational, 10 computational only), bringing the total requested time to 320,000 hours (0.25 hours x 64 cores x 1000 structures x 20 runs). 2. The structure of CdTe grain-boundaries from STEM images. We will consider large scale experimental-computational optimization with a periodic GA using LAMMPS and a STEM image similarity code, developed by a member of our group, to optimize the energy and comparison to experimental data. We also will optimize smaller scale grain-boundaries using the DFT code VASP and the same STEM image similarity code. Since the exact number of atoms in the grain-boundary is undetermined, we will run 20 different calculations with varying ratios of Cd:Te. Each geometry optimization will take 30 hours. These calculations have a two-fold importance: to develop our GA code to treat periodic boundary conditions, and to determine the structure of experimental grain-boundaries. (2 calculation types x 20 calculations x 30 hours x 256 core hours=307,200 hours) 3. The structural optimization of solid interfaces in Li-ion batteries and Si photovoltaics. These structures will be determined using the same GA code with periodic boundary conditions as in #2 above, and using DFT calculations with VASP. Running 10 optimizations for each of 5 interfaces, we expect each run to run to require 15 hours on 256 cores. (2 materials x 50 calculations x 15 hours x 256 core hours=384,000 hours) The total production run hours above add up to 1,011,200 hours. We estimate using roughly an additional 190,000 core hours for debugging, benchmarking and convergence tests. We require the following packages to be installed: • Anaconda Python 2.7 • VASP 5.4.4 Industry partnership: N/A Project URL: Requested allocation: 1200000 Q1: 300000 Q2: 300000 Q3: 300000 Q4: 300000 Justification: All systems considered have been tested on KNL nodes on Cori at NERSC. As such, the required time for each previous result is relevant, since both Cori and Bebop use KNL nodes. The DFT calculations mentioned above have been benchmarked on Knight’s Landing nodes (KNL). The parallel performance of our genetic algorithm code has been tested on Au nanocluster test cases. When 1 KNL node is used, performing 4 VASP calculations, each model takes 41.76 seconds. If only one VASP calculation is run on 1 KNL node, it takes 29.69 seconds. This means, that by running 4 VASP calculations simultaneously on a KNL node, about 3x the total number of VASP calculations can be run per node-hour compared to only 1 calculation at a time. For the second and third systems considered, previous results from Cori KNL VASP calculations show that there is a weak-scaling parallel performance. When 4 KNL nodes are used to run VASP calculations there is a parallel efficiency of 106% when compared to a calculation on 2 KNL nodes. Storage requirements: 1 TB The requester has used undetermined amount hours of their initial startup project. In addition to approving an initial amount, please specify a Category and Subcategory for this project. For a list of the current selection of approved categories, please see: https://wiki.lcrc.anl.gov/wiki/Processes/Categories Once the Allocation committee has approved the project, please go to the Project Management page to create it: https://accounts.lcrc.anl.gov/projects.php Thank You, The LCRC Accounts System