Parallel optimization for model calibration of large scale particle systems
Typ
Examensarbete för masterexamen
Master's Thesis
Master's Thesis
Program
Engineering mathematics and computational science (MPENM), MSc
Publicerad
2023
Författare
Samuelsson, Alexander
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
In this thesis we present an optimization framework for model calibration of Discrete
Element Method (DEM) simulations. The DEM is a method for simulating large
scale particle systems. Numerical complexity makes calibration of models computa tionally expensive and time consuming. We approach the procedure by viewing it
as a black box optimization problem and apply a parallel optimization algorithm to
perform calibration.
The algorithm is a surrogate optimization method using Radial Basis Functions.
We first sample the objective function with a Symmetric Latin Hypercube serving
as a design of experiments. In the iterative part of the algorithm we implement the
Metric Stochastic Response surface method proposed by Regis and Shoemaker. We
choose new evaluation points based on a weighted score of the expected objective
value of the point and its distance from previously evaluated points. The weight
decides whether to focus on exploitation by trying to minimize the objective value of
the next point, or exploration by evaluating points in areas of the sample space where
we have less information. To reduce the wall time required in finding a solution the
framework allows for asynchronous parallelism by simply evaluating many points at
the same time. We also implement a method for early termination.
The algorithm is applied to two DEM optimization problems. The first one is a
calibration problem involving a device used to calibrate the friction of a material.
The second one is an industrial optimization case where the position of a funnel is
adjusted to achieve an even split of material flow. In both cases, the algorithm con verges to a solution in less than 40 evaluations. Furthermore the early termination
and parallel strategies are shown to reduce the time required to solve the calibration
case. We finish by discussing and comparing our results, giving some advantages
and disadvantages found with the early termination functionality.
Beskrivning
Ämne/nyckelord
DEM, Discrete element method, Calibration, Optimization, Surrogate, Black-box, Parallelism