Performance of Neural Networks
Brief explanation of neural networks Four classes of innovations based on neural networks Are papers on neural networks respectable? Materials Science
Empirical Equations
noise
modelling uncertainty
elegantly answers the question as to whether sufficient data have been used to create the model unreasonable to ask if sufficient data used to create model, relative to number of coefficients - may be sufficient data in some regions of input space and not in others
International Fusion Reactor Reduced activation steels
Journal of Nuclear Materials 348 (2006) Kemp, Cottrell & Bhadeshia
Exploitation of neural networks Discover new science Explain observations Design materials or processes Quantitative expression of data
Neural networks: unexpected outcomes
tested at room temperature 14.3% 7%
tested at 100 °C 14.3% 13%
Neural networks: design
Keehan, Karlsson, Andrén, Bhadeshia, Science & Techn. Welding & Joining 11 (2006) Ni 2Mn
Exploitation of neural networks Discover new science Explain observations Design materials or processes Quantitative expression of data
9Cr1Mo Dimitriu & Bhadeshia, 2007
9Cr1Mo Dimitriu & Bhadeshia, 2007
precipitates solid solution iron + microstructure 550 °C 600 °C Murugananth & Bhadeshia, 2001 Components of Creep Strength 2.25Cr1Mo
Exploitation of neural networks Discover new science Explain observations Design materials, processes, experiments Quantitative expression of data
Suppose we fail to achieve 650°C Ferritic Creep-Resistant Steel
Exploitation of neural networks Discover new science Explain observations Design materials, processes, experiments Quantitative expression of data
weld pool shape Mishra and DebRoy, MSE A, 454 (2007) 454
Five steps in the creation of meaningful neural networks
Creation of model Make predictions using model Modelling uncertaintyInvestigate prediction experimentally Model or data disseminated
Creation of model1 Make predictions using model2 Investigate prediction experimentally2 Modelling uncertainty2 Model or data disseminated3