You are on page 1of 3

1

Using the Global Optimization Toolbox As an


Optimization Tool For Calibrations

Jonathan Ye
08/03/22
2

Introduction
In this document, I will discuss using the Global Optimization Toolbox to help the
IPSD-L Mechatronics team tune some Selective Catalytic Reduction (SCR) models that depicts
several chemical reactions occurring during aftertreatment. Briefly, the SCR uses ammonia and
catalysts to turn harmful NOx into nitrogen gas. More specifically, these tunes describe the
kinetics of the chemical reactions including absorption and desorption of the ammonia binded
catalyst, a standard reaction to create N2, a slow reaction to create N2, and a fast reaction of N2.
In a previous document, ‘Using Fmincon As an Optimization Tool For Calibrations’, I discussed
how to use ‘fmincon’ , a gradient descent based optimizer. In this document I will explain the
shortcomings of a gradient descent based optimizer for more non-smooth objective functions
with multiple local minimums.
Background
Previously, with the HC dosing FF model, a gradient-descent optimization function from
MATlab’s optimization toolbox (‘fmincon’). However, after trying to adapt that same function
for the SCR kinetics tunes, it seems that the objective function (root mean squared error) has
multiple local minimums, rendering the gradient-descent based optimization function, much less
effective. So, we need a more global search for a minimum of the objective function, which has
turned us to using the Global Optimization Toolbox that MathWorks provides. In this document,
I will go over using ‘patternsearch’ which is a good starting global optimization function [1].
For guidance on what data we used for which reaction, the following list can be used:

Data - Umicore Gen 4 Core Bench Data for Simapp Model - OneDrive (sharepoint.com) (Sent
by Lauren Duvall, so contact her if there are any questions/issues)

1. Absorption and desorption of the catalyst → Degreened/200211 Umicore CuSCR M500


Gen 4_NH3 ads_des NH3_3000_R4_500ppm.dat
2. Standard reaction → Degreened/ NH3 300 method data/ 200224 Umicore Gen 4
M500_4steps experiment_ANR Sweep 500 ppm.xlsx AND Degreened/ NH3 300 method
data/ 200224 Umicore Gen 4 M500_4steps experiment_ANR Sweep 1000 ppm.xlsx
3. Slow reaction → Degreened/ NH3 300 method data/ 200224 Umicore Gen 4
M500_4steps experiment_ANR Sweep 1000 ppm.xlsx
4. Fast reaction → Degreened/ NH3 300 method data/ 200224 Umicore Gen 4 M500_4steps
experiment_ANR Sweep 1000 ppm.xlsx
5. Total reaction validation → Degreened/ NH3 300 method data/ 200204 Umicore Gen
4_5step experiment.xlsx

The specific data channels that were used can be found in each respective excel sheet. It
is important to know that the following preprocessing of the data was done of each data set:
3

1. Absorption and desorption of the catalyst → Got rid of steady state values (this was done
through the Simulink scope; hovered over desired time values and processed them out of
the data in MATlab)
2. Standard reaction → Kept all data
3. Slow reaction → Got rid of all data above 420 ℃ and made sure to only use the data
where NO2/Nox >= 50% (Done through MATlab script)
4. Fast reaction → Got rid of all data above 420 ℃ and made sure to only use the data
where NO2/Nox < 50% (Done through MATlab script)

For a lot of good information, refer to the reference link. It talks a lot about the Global
Optimization Toolbox and how to use it.

References
[1]https://www.researchgate.net/profile/Olli-Tahvonen/publication/262956005_On_the_economi
cs_of_optimal_timber_production_in_boreal_Scots_pine_stands/links/573068a208ae744151911
731/On-the-economics-of-optimal-timber-production-in-boreal-Scots-pine-stands.pdf

You might also like