Bayesian Neural Networks and Irradiated Materials Properties - PowerPoint PPT Presentation

1 / 33
About This Presentation
Title:

Bayesian Neural Networks and Irradiated Materials Properties

Description:

Bayesian Neural Networks and Irradiated Materials Properties. Richard Kemp ... z = 0.8[tanh(nx-2) tanh(x2-n) tanh(ny 2) tanh(y2-n) 1] ... – PowerPoint PPT presentation

Number of Views:57
Avg rating:3.0/5.0
Slides: 34
Provided by: richar403
Category:

less

Transcript and Presenter's Notes

Title: Bayesian Neural Networks and Irradiated Materials Properties


1
Bayesian Neural Networks and Irradiated Materials
Properties
  • Richard Kemp
  • University of Cambridge

2
  • Neural networks
  • (and why Bayes?)
  • Modelling materials properties
  • Genetic algorithms
  • Materials Algorithm Project (MAP)

3
Problems
  • Prediction of irradiation hardening
  • Prediction of irradiation embrittlement
  • Physical models?

4
A simple neural network
5
A simple neural network
z 0.8tanh(nx-2) tanh(x2-n) tanh(ny2)
tanh(y2-n) 1
(i.e. two inputs and four hidden units)
6
Why Bayes?
Predict the next two numbers 2, 4, 6, 8 ?
7
(No Transcript)
8
(No Transcript)
9
Bayesian neural networks
10
ANN design
  • Data availability
  • Dimensionality reduction?
  • Over/under fitting

11
Fitting error
(Number of hidden units)
12
materials modelling
13
Modelling irradiation hardening
  • No current strongly predictive model
  • Data collected by Yamamoto et al and from
    European RAFM database
  • 1800 data up to 90 dpa
  • 36 input variables
  • No heat treatment information included

14
Inhomogeneous data
15
Testing of physics
  • Saturation?
  • Arrhenius (temperature-dependent) effects?
  • Helium effects?

16
(No Transcript)
17
Model performance
18
Model performance
Unirradiated Eurofer 97
19
Model performance
Unirradiated and irradiated F82H
20
Modelling irradiation embrittlement
  • Modelling Charpy ?DBTT
  • Miniaturised specimens for fusion materials
    research
  • 461 data available
  • 26 input variables
  • Heat treatment data included
  • Reduced compositional information

21
(No Transcript)
22
(No Transcript)
23
(No Transcript)
24
Effects of chromium
25
Effects of phosphorus
26
Extrapolation to fusion?
Eurofer 97 yield stress
27
Genetic algorithms
28
Circle of life
29
Genetic algorithms
  • Cope with non-linear functions
  • Cope with large numbers of variables efficiently
  • Cope with modelling uncertainties
  • Do not require knowledge of the function

30
0.13C-9Cr-2W-0.1Ta-0.15V-0.25Mn
31
Further issues
  • Missing data
  • Confounding factors and correlations
  • Fusion-relevant irradiation?
  • Genetic algorithm design
  • Satisfaction of multiple design criteria

32
(No Transcript)
33
  • Thanks to Geoff Cottrell and Harry Bhadeshia

www.msm.cam.ac.uk/phase-trans
Write a Comment
User Comments (0)
About PowerShow.com