A neural network approach for solution of the inverse problem for selection of powder metallurgy materials

This paper describes research that has been conducted into artificial intelligence techniques for solving the ‘inverse problem’, for assisting with materials selection. The term inverse problem refers to the task of employing process output information (i.e. the required mechanical or physical properties of the final material), in order to recommend suitable input settings for the process concerned. For example, for the powder metallurgy (P/M) process, where parts are manufactured from powdered metals, powder composition, compaction pressure, and sintering conditions are important input parameters that have to be controlled. Previous attempts at solution of the inverse problem have involved the use of statistical methods (such as regression analysis with application of relevant transforms), for fitting curves to the available experimental data. The resulting equations can be combined in a rule-base for generating materials selection advice. While such techniques are useful for identifying general trends in process inputs and outputs, they are subject to a number of disadvantages. P/M manufacture involves multiple process inputs and outputs. Many of the relationships are non-linear, and the experimental data exhibits considerable noise. When fitting curves to non-linear data the selection of transforms is inevitably subjective and becomes very difficult when multiple inputs are involved; also, regression analysis is not well suited to modelling noisy data. These considerations have lead to the identification of the neural network approach as being suitable for P/M modelling for materials selection purposes. Multiple inputs, modelling of highly non-linear responses, and the avoidance of detrimental noise effects have been provided by training a backpropagation neural network with experimental data for ferrous P/M data. The neural network deductions for process inputs were compared to those generated by regression analysis. The network reduced the standard deviation of the errors associated with the inverse solutions by 36%, thereby demonstrating how the technique can improve the accuracy of process recommendations.