1. Application of artificial neural network in materials research
Wei SHA
Professor of Materials Science
25 June 2007

2. Modelling methodologies Metals Research Group, Queen’s University
Thermodynamic modelling
The Johnson-Mehl-Avrami method and its adaption to continuous cooling and heating
Finite element method
Phase field method
Atomistic simulation
Neural network method

3. The models Integrated
Malinov, Sha, JOM, 57(11), 2005, 54.

4. The models Graphical user interfaces of software for modelling
Composition-processing-temperature-mechanical properties TTT diagrams
Fatigue life CCT diagrams Microhardness profile
Malinov, Sha, Computational Materials Science, 28, 2003, 179.

5. Basic principles of neural network modelling Natual Artificial
Neuron
Input Layer
Hidden Layer
Output Layer
Neural Network
The human brain contains
1010 – 1011 neurons
Malinov, Sha, McKeown, Computational Materials Science, 21, 2001, 375.

6. Basic principles of neural network modelling Architecture of the neural network
Malinov, Sha, Guo, Materials Science and Engineering A, 283, 2000, 1.

7. Basic principles of neural network modelling Neural network modelling steps
Database collection
Analysis and pre-processing of the data
Design and training of the neural network
Test of the trained network
Using the trained NN for simulation and prediction

8. Database construction and analysis Training database

9. Database construction and analysis Distribution of the input dataset
Guo, Malinov, Sha, Computational Materials Science, 32, 2005, 1.

10. Algorithm of computer program Creation of neural network model
Reading of file with database
Normalisation of the data
Creating neural network and defining training parameters
Neural network training
Post-training analyses for training and test subsets
Use of the model
Experimental verification
Random redistribution of database
Dividing to training and test subsets for inputs and corresponding outputs
Loop for new random redistribution of the database
Loops for new network architecture, training algorithm, transfer function and training parameters

11. Algorithm of computer program Steps in creating the model
Master database containing all the data for each input and output
Sub-spreadsheets for each individual output
Matrices in sub-spreadsheets copied-and-pasted into a notepad file for each output
MATLAB files created for the training of each individual neural network
Individual neural networks ‘trained’ to best possible accuracy, resulting data files saved for each output
Master model file created using MATLAB to read in the output mat files from the training, and predict the outputs. Model runs from a central GUI (Graphical User Interface) for user control
McBride, Malinov, Sha, Materials Science and Engineering A, 384, 2004, 129.

12. Neural Network Predictions
Algorithm of computer program Post-training linear regression analysis
500
1000
1500
2000
2500
Training
R = 0.97
Neural Network Predictions
Experimental
Test
R = 0.963
Data Points
A = T
Best Linear Fit

13. Algorithm of computer program Training parameters

14. Algorithm of computer program Post-training validation of the software simulations
400
Min =
Max = 35.89
Mean = -0.13
STDEV = 10.31
350
300
250
Number
200
150
100 50
-40
-30
-20
-10 10 20 30 40
×100 (%) HV
Error
EXP NN - =

15. Algorithm of computer program Comparison between prediction and experiments
Sha, JOM, 58(9), 2006, 64.

16. Noise tolerance The experimental values are values including noise

17. Noise tolerance The experimental values are values without noise

18. Use of the software Block diagram of software system for modelling
Databases
Computer program for training (learning)
Trained artificial neural networks
Graphical User Interfaces for use of the models
Graphical User Interfaces for upgrade of the models
Module for new data input
Module for input of re-training parameters
Module for materials selection

19. Use of the software Influence of alloy composition in g-TiAl, 1040 °C
Malinov, Sha, Materials Science and Engineering A, 365, 2004, 202.

20. Use of the software Microhardness profiles of titanium after nitriding
Zhecheva, Malinov, Sha, JOM, 59(6), 2007, 38.

21. Use of the software Ti–15Mo–5Zr–3Al, nitrided in N2 at 750 °C for 60 h
Zhecheva, Malinov, Sha, Surface and Coatings Technology, 200, 2005, 2332.

22. Use of the software Optimization of the alloy composition & processing
Optimisation
Criteria
Trained
Neural
Network
Solution
Loops for
Heat Treatment
Temperature
Alloy Composition
Find
Alloy composition with
max strength at 420°C
Fix
Heat treatment = Annealing
T = 420°C
Sn, Cr, Fe, Si, Nb, Mn = 0; O = 0.12
Vary
Al, Mo, Zr, V
Al = 5.8; Mo = 7.3; Zr = 5.2; V = 0
Tensile strength (420°C) = 932 MPa;
Yield strength = 665 MPa;
Elongation = 10%;
Modulus of elasticity = 94 GPa;
Fatigue strength = 448 MPa;
Fracture toughness = 101 MPa m1/2