Neural Networks

1
Neural Networks
www.msm.cam.ac.uk/phase-trans
2
(No Transcript)
3

4
(No Transcript)
5
There are many problems where simplification is
unacceptable
6
(No Transcript)
7
(No Transcript)
8
Charpy
fatigue
tensile
critical stress intensity
corrosion
9

• Given a comprehensive description of material,
  process and structure, it is not yet possible to
  predict most properties.

10
Variables

• C, Mn, Si, Ni, Cr, Mo, V, Co, B, N, O..
• Thermomechanical processing of steel
• Welding consumable
• Welding parameters
• Subsequent heat treatment

11
y
x
Michael McIntyre
12
Solution

• non-linear functions
• large numbers of variables
• uncertainties
• exploit large knowledge base

13
Empirical Equations
?y a b (C) c (Mn) d (Ni)
....
14
?y a b (C) c (Mn) ?y a b (C) c
(Mn) d(C x Mn)
15
?y a b (C) c (Mn) ?y a b (C) c
(Mn) d(C x Mn) ?y sin (C)
tanh (Mn)
16
(No Transcript)
17
Hyperbolic Tangents
18
(No Transcript)
19
(No Transcript)
20
(No Transcript)
21
(No Transcript)
22
(No Transcript)
23
(No Transcript)
24
(No Transcript)
25
(No Transcript)
26
y
A
B
x
27
(No Transcript)
28
What is the range over which an empirical method
should be used?
29
Outliers
30
Cole Bhadeshia, 1999
31
GTA weld at 823 K (data from Nippon Steel)
600
500
400
300
200
100
0
20000
30000
40000
Life / hours
Cole Bhadeshia, 1999
32
(No Transcript)
33
Cool, 1996
34
Cool, 1996
35
600 C
As-welded
700 C
650 C
Cool, 1996
36
Siemens Mitsui Babcock Nippon Steel ABB
37
Murugananth Bhadeshia, 2002
38
Coalesced bainite
Keehan, Karlsson, Andrén and Bhadeshia, 2005
39
Nickel base alloy FT750dc
wt
40
(No Transcript)
41
Tancret Bhadeshia, 2002
1000
800
600
Yield stress / MPa
400
200
0
0
200
400
600
800
1000
1200
Temperature / C
42
International Fusion Reactor
Reduced activation steels
43
Fusion Reactor Steels Kemp, Cottrell Bhadeshia,
2006
44
(No Transcript)
45
Thank you
46
Classification networks
47
(No Transcript)
48
(No Transcript)
49
(No Transcript)
50
(No Transcript)