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: Fatih Sen Applicant's institution: ANL Applicant's division: NST Project Name: FF-opt Project title: Comparison of optimization algorithms for force field fitting Associated funding: LDRD Other Systems: Science: Accurate description of atomic processes such as catalytic activity or reaction kinetics with empirical force fields strongly depends on the force field parameterization process. An empirical force field gives the total energy (E) of a system of atoms based on the positions of each atom, i.e. E=f({a_i},{r_n}), where f is a function that describes the interaction, {a_i} represents the set of parameters, and {r_n} the set of atomic positions. The determination of force-field parameters {a_i} for various complex functional forms requires a systematic approach, which can enable sampling of most of the available parameter space. In this project, we will explore different local and global optimization algorithms in the optimization of force field parameters including single and multi-objective genetic algorithms, and compare the effectiveness of different algorithms in the parameterizing of different functional forms. In terms of local optimization, we wil l test the commonly used Simplex and Levenberg-Marquardt (L-M) methods that are known to perform reasonably well for many non-linear least-squares problems. In addition to these common methods, in collaboration with Wild Group at MCS, we will use a more sophisticated derivative-free method for non-linear least squares problems as implemented in POUNDERS and ORBIT. Local optimization methods rely greatly on the initial guess for the unknown parameters and fails for very complex functional forms. To cover all of the parameter phase space, a large number of starting guesses needs to be evaluated. On the other hand, global optimization methods comprehensively search suitable regions of parameter space and can be more suitable for force-field parameterization. Here, we will employ evolutionary algorithms namely genetic algorithms (GA) in force field fitting, which are reported to exhibit unbiased, and a more global search by maintaining a diverse population, hence discovering potentially good regions of interest. In all local and global optimization schemes, the goodness of fit and hence, the predictive capability of the force field, depends on the weighting scheme selected between different observables. The selection of weights is typically based on the difference in the numerical values and total number of different quantities, and also based on the subjective determination related to the problem of interest. One way of removing the weighting effects is to use a multi-objective optimization algorithm, which can give a set of solutions at the Pareto front and enable the selection of force field parameters according to desired properties. We will explore the applicability of multi-objective GA for determining force field parameters and compare the results with single-objective GA. Project description: The comparison of different optimization algorithms will be carried out on using an existing training dataset obtained from density functional theory calculations (DFT) using the VASP code, which contains binding energies, elastic constants, lattice constants and internal coordinates for various IrO2 polymorphs. We are planning to explore 3 different functional forms (f), namely Morse, Morse+QEq and Tersoff, as implemented in the GULP software. The test of Simplex and L-M will be done using the program codes implemented in Numerical Recipes and MINPACK. These codes essentially work in serial, but we will employ multi-start algorithms to run different initial guesses in parallel. The POUNDERS and ORBIT algorithms will be tested as implemented in the PETSc code available in LCRC resources. For single- and multi-objective genetic algorithms, we will use the PyGMO program, which is already installed in the Blues system. PyGMO can be efficiently (100%) pa rallelized over the population size used. The details of the calculations can be listed as: i) We will compare different optimization techniques for force field parameters with different functional forms. We are planning to use 500 initial guess parameter set for the Simplex and L-M algorithms. For Simplex, each optimization calculation takes 4 hours in 16 cores. For L-M, each optimization calculation takes 2 hours in 16 cores. So, in total we will require 3*500*(16*2+16*4) = 144,000 core hours. ii) In single objective GA tests, we are going to investigate effect of population size (4 different population size of 100, 200, 400, and 1000) on the ability of the potential parameters to reproduce DFT results, and the speed of convergence. On average, each GA run is estimated to require 60 hours on 128 cores. So the calculations will require 2*3*4*60*128= 184,320 core hours. ii) In multi-objective GA tests, we are going to investigate effect of population size (4 different population size of 100, 200, 400, and 1000). On average, each GA run is estimated to require 60 hours on 128 cores. This part of the project will therefore require 2*3*4*60*128= 92,160 core hours. Total computation time requested: 420,000 core hours. Industry partnership: Project URL: Requested allocation: 420000 Q1: 0 Q2: 0 Q3: 0 Q4: 420000 Justification: Storage requirements: 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