optimal design : surrogate models
TRANSCRIPT
![Page 1: Optimal Design : Surrogate Models](https://reader036.vdocument.in/reader036/viewer/2022081620/62ceec175e19683f1e5f10a3/html5/thumbnails/1.jpg)
Optimal Design : Surrogate Models
Kuei-Yuan ChanNational Taiwan University
1
![Page 2: Optimal Design : Surrogate Models](https://reader036.vdocument.in/reader036/viewer/2022081620/62ceec175e19683f1e5f10a3/html5/thumbnails/2.jpg)
A good model should represent reality in the simplest meaningful manner.
A surrogate model is a simpler analysis model extracted from the more sophisticated ones using a variety of data-handling techniques.
2
![Page 3: Optimal Design : Surrogate Models](https://reader036.vdocument.in/reader036/viewer/2022081620/62ceec175e19683f1e5f10a3/html5/thumbnails/3.jpg)
Model forms
• Analytical models in mathematical form : based on fundamental science with proper assumptions. Generally easier to calculate but with limited applications.
• Computer simulations : most commonly used in engineering practice. These codes are wrapped within a graphical user interface (GUI). One must understand the limitations and assumptions behind these computer logics.
• Experiment data fitting : also one of the most common practice in engineering. Although in math forms, it does not provide basic science and usually limited to interpolations.
3
![Page 4: Optimal Design : Surrogate Models](https://reader036.vdocument.in/reader036/viewer/2022081620/62ceec175e19683f1e5f10a3/html5/thumbnails/4.jpg)
Wind Tunnel with Fluid Dynamics Modeling
4
credit to 陳亮宇、沈育儒、柯曼德、楊敦仁、徐征宇
![Page 5: Optimal Design : Surrogate Models](https://reader036.vdocument.in/reader036/viewer/2022081620/62ceec175e19683f1e5f10a3/html5/thumbnails/5.jpg)
Curve fitting with various models
5
0
5
10
15
20
0 3 6 9 12
![Page 6: Optimal Design : Surrogate Models](https://reader036.vdocument.in/reader036/viewer/2022081620/62ceec175e19683f1e5f10a3/html5/thumbnails/6.jpg)
Linear and Quadratic Fits and Least Squares
• Let be the approximation function.
• The least square best linear fit is the function such that
6
Expensive
x1
x2
...xm
y1
y2
...ym
y
mina0,a1
�m⇤
i=1
yi � yi
⇥2
subject to yi = aT1 xi + a0
![Page 7: Optimal Design : Surrogate Models](https://reader036.vdocument.in/reader036/viewer/2022081620/62ceec175e19683f1e5f10a3/html5/thumbnails/7.jpg)
Neural Network Modeling
7
![Page 8: Optimal Design : Surrogate Models](https://reader036.vdocument.in/reader036/viewer/2022081620/62ceec175e19683f1e5f10a3/html5/thumbnails/8.jpg)
Basic Neuron
• Basic neuron without bias
• Basic neuron with bias
8
![Page 9: Optimal Design : Surrogate Models](https://reader036.vdocument.in/reader036/viewer/2022081620/62ceec175e19683f1e5f10a3/html5/thumbnails/9.jpg)
General Neural Network Model
9
y(x) = �
�n⇤
i=1
wixi + b
⇥weights
activation function
![Page 10: Optimal Design : Surrogate Models](https://reader036.vdocument.in/reader036/viewer/2022081620/62ceec175e19683f1e5f10a3/html5/thumbnails/10.jpg)
Kriging Modeling
10
![Page 11: Optimal Design : Surrogate Models](https://reader036.vdocument.in/reader036/viewer/2022081620/62ceec175e19683f1e5f10a3/html5/thumbnails/11.jpg)
Kriging Model
11
y(x) = f(x) + z(x)
f(x) = � = (IT R�1I)�1IT R�1yR(xi,xj) = e
�
n�
k=1
�k|xik � xj
k|2
� = argmin�|R| 1
m ⇥2⇥
⇥2 =1m
(y � �I)R�1(y � �I)
is in the form
where
with
![Page 12: Optimal Design : Surrogate Models](https://reader036.vdocument.in/reader036/viewer/2022081620/62ceec175e19683f1e5f10a3/html5/thumbnails/12.jpg)
Matlab Functions
• NEWFF
• TRAIN
• kriging demo.zip
• kriging-ych.zip
12