Radu Calin Dimitriu Modelling the Hot-Strength of Creep-Resistant Ferritic Steels and Relationship to Creep-Rupture Data
Outline: - 1 st Part The Technique - 2 nd Part The Model
The Technique Neural networks in a Bayesian framework was employed to model the residual stress A neural network is a non-linear method of regression In neural networks data are fitted to a function in order to capture the complex relations that exist between the inputs and the output
Structure of a three layer neural network
Input variables C, Mn, Si, Ni, Cr, Mo, V, Co, B, N, O, … Thermomechanical processing of steel Subsequent heat treatment Steel properties Operating conditions...
The general form of the neural network y – represents the output w – represents the weights h – represents the hidden units θ – represents the bias x – represents the inputs i, j – represent subscripts
Hyperbolic Tangents
Problem Powerful non-linear regression methods → overfitting
Training and testing data sets
Advantage of neural networks in a Bayesian framework - the uncertainties are quantified - ant then transformed in error bars
Outline: - 1 st Part The Technique - 2 nd Part The Model
Significance
Predictions
Calculations
Correlation
ACKNOWLEDGEMENTS Many thanks to professor H.K.D.H. Bhadeshia The Department of Material Science and Metallurgy The University of Cambridge The European Commission under the Marie Curie Early Stage Research Training Programme