Artificial neural network modelling is a non-linear statistical analysis technique that links input data to output data using a particular set of non-linear functions. Especially for alloys which have no explicit physical model describing qualitatively the relationships between alloy composition, processing parameters and the final properties, an artificial neural network maximises the output of combinatorial research.
Artificial neural network modelling provides a way of using examples of a target function to find the coefficients that make a certain mapping function approximate the target function as closely as possible.
Understanding the correlations between alloy composition, processing parameters, microstructures and their final properties is of great importance for alloy development, since these relations govern the alloy design and production.
Especially for alloys which have no explicit physical model describing qualitatively the relationships between alloy composition, processing parameters and the final properties, a model based on an artificial neural network can easily be created with the existing data. For this, very little or no prior knowledge of the physical background of the relationships is needed. This is perfect for exploratory research.
OCAS developed a model to predict the hardness of new alloys based on artificial neural network approach. This model will support the development of future alloys by optimising composition design and processing. Based only on the input parameters – the concentration of 13 elements- and the aging treatment (temperature and time), the model is able to predict the hardness of these alloys.
Model performance & assessment
Model performance: the performance of the model (comparison between the actual and the predicted hardness) shows a regression coefficient >0.9
To experimentally check the performance of the artificial neural network model, 20 new compositions were processed. Two groups of compositions were selected alloys with compositions close to the ones on the database and alloys with additions of other elements or outside the database range.
The hardness predicted and the experimentally evaluated is displayed in the figure below. The alloys with a composition close to the ones in the database show a good correlation between the calculated hardness and the experimental value (green circle). This supports the use of the artificial neural network as a powerful tool to fine tune the chemical compositions and heat treatments for compositions close to the ones in the database. On the other hand, the figure evidences the limits of the artificial neural network model when predicting the hardness for alloys out of the database ranges.
“The use of artificial neural network models truly maximises the output of combinatorial research”