Iqbal Shah and Thomas
Phase Transformations Group,
Department of Materials Science and Metallurgy,
University of Cambridge,
Cambridge CB2 3QZ, U.K.
The neural network program was produced by:
University of Cambridge,
Cambridge, CB3 0HE, U.K.
Added to MAP: September 2002.
To estimate the yield strength of austenitic stainless steels as a function of temperature, chemical composition and heat treatments.
FORTRAN / C
Source code / Executable files
Solaris 5.5.1 & Windows 95
The modelling procedure is a purely empirical one, and is based on a neural network program called generate44, which was developed by David MacKay and is part of the bigback5 program. The model is constituted of a committee of several individual neural networks. It was trained on a set of experimental data for which the "outputs" are known, and creates a kind of non-linear, multi-parameter "regression" of the outputs versus the inputs. This "regression" has already been produced and the model is delivered ready to perform predictions for steels of any desired composition (within certain specified limits). The source code for the neural network program can be downloaded from David MacKay's website; the executable files only are available from MAP.
The .tar file contains the following files:
A manual containing step-by-step instructions for running the program, including a list of input variables.
A text file containing the minimum and maximum limits of each input and output variable. This file is used to normalise and unnormalise the input and output data.
An input text file containing the input variables used for predictions, together with an example set of data.
Contains the results you should expect from the example set of data. To test the model is running properly on your computer, use the given 'test.dat' file to do predictions and compare the 'result' file with this file.
This is a unix shell file containing the command steps required to run the module. It can be executed by typing sh model.gen at the command prompt. These shell files run all the programs necessary for normalising the input data, executing the network for each model, unnormalising the output data and combining the results of each model to produce the final committee result.
Created by generate_spec, which contains information about the module and the number of data items being supplied. It is read by the program generate44.
This executable file creates a file called spec.t1, required by generate44.
This executable file creates a file called norm_test.in, which contains the normalised equivalent of the input data found in test.dat. It requires the MINMAX file
This is the executable file for the neural network program. It reads the normalised input data file, norm_test.in (created by normalise) , and uses the weight files in subdirectory c, to find a value for the output. The results are written to the temporary output file _out.
This executable file combines the predictions of the different models in the committee and calculates the combined error bar. .treatout
This executable un-normalise the committee predictions and produces the file 'result'.
Contains the final un-normalised committee results for the predicted output.
The weights files for the different models.
Files containing information for calculating the size of the error bars for the different models.
Files containing information about the perceived significance value for each model.
Files containing values for the noise, test error and log predictive error for each model.
A normalised output file which was created during the building of the model. It is accessed by generate44 via spec.t1.
The normalised output file containing the committee results. It is generated by .gencom.
Detailed instructions on the use of the program are given in the README file.
The input variables for the model are listed in the README or README.DOC file in the corresponding directory. The maximum and minimum values for each variable are given in the file MINMAX.
These programs give the yield strength in 'MPa'. The corresponding output files is called Model_RESULT.dat or Result. The format of the output file is:
Prediction Error Upper-limit Lower-limit
(MPa) (MPa) (MPa)
An estimated predictive error bar is provided by the model.
See Neuromat Ltd for a user friendly way of making predictions.
file in a dedicated directory (for example: "neural").
On UNIX systems, this is done by:
For Solaris 5.5.1, Linux or SGI, just type:
The results are written in the "Result" or "model_result.dat" file, as described in the README file. In the present case:
neural network, yield strength, austenitic stainless steel
MAP originated from a joint project of the National Physical Laboratory and the University of Cambridge.MAP Website administration / firstname.lastname@example.org