Keywords

Gaussian Process, Bayesian Global Optimization, Limited Data-budget, GUI

Presentation Type

Event

Research Abstract

Design-optimization with realistic computer codes is a ubiquitously challenging task. Typically, we have to execute thousands of simulations in order to achieve a globally optimum design. However, since realistic models may take hours or even days to complete a single simulation, global optimization is infeasible for all but the simplest models. We are necessarily limited to just a handful of simulations. Bayesian global optimization (BGO) is a computational framework built upon Gaussian process regression that allows us to actively select which simulations to make in order to reach our objective or its gradients. It only assumes that the objective is measurable at any given design point either experimentally or via a computer simulation. We have implemented BGO in Python and created a nanoHUB tool that applies the concept to the problem of the structure determination of an arbitrary cluster of atoms. The tool works as follows: First, it generates an initial data pool consisting of random structures and their associated energies as well as a test design pool consisting of structures that will be tested for optimality. Then, it constructs a Gaussian process model of the energy surface and employs BGO to find the minimum energy cluster among the test pool. The process runs until either the maximum expected improvement of future simulations falls below a threshold or the number of maximum iterations is reached.

Session Track

Sustainability

Share

COinS
 
Aug 6th, 12:00 AM

Design-Optimization with a Limited Data-Budget

Design-optimization with realistic computer codes is a ubiquitously challenging task. Typically, we have to execute thousands of simulations in order to achieve a globally optimum design. However, since realistic models may take hours or even days to complete a single simulation, global optimization is infeasible for all but the simplest models. We are necessarily limited to just a handful of simulations. Bayesian global optimization (BGO) is a computational framework built upon Gaussian process regression that allows us to actively select which simulations to make in order to reach our objective or its gradients. It only assumes that the objective is measurable at any given design point either experimentally or via a computer simulation. We have implemented BGO in Python and created a nanoHUB tool that applies the concept to the problem of the structure determination of an arbitrary cluster of atoms. The tool works as follows: First, it generates an initial data pool consisting of random structures and their associated energies as well as a test design pool consisting of structures that will be tested for optimality. Then, it constructs a Gaussian process model of the energy surface and employs BGO to find the minimum energy cluster among the test pool. The process runs until either the maximum expected improvement of future simulations falls below a threshold or the number of maximum iterations is reached.