The goal of this project is to be able to predict the changes which occur in the grain structure of aluminium and nickel superalloy components when they undergo a forging process. Forging is a method of deforming a metal at high temperatures into a component shape. This deformation process and any subsequent annealing stage controls the microstructure (grain size, texture, etc.) of the component and has a dominant effect on the final macroscopic properties of the material.
At present, no reliable model exists for determining the final microstructure. Thus proper optimisation of the forging process normally demands a combination of experience and expensive trials. Because the final microstructure is dependent on many parameters (alloy composition, working temperature, local strain rate, etc.) it is very difficult to model with traditional micromechanical models. Neural networks, however, are ideally suited to such non-linear, multi-parameter problems. Part of my research is to investigate and develop suitable neural network architectures implementing Bayesian methods which are appropriate for this metallurgical application. Such an approach will not only provide a predictive capability for the process but will also provide a rigorous assessment of the reliability of these predictions. This is both a very interesting and important project, because it lies at the boundary between practical industrial problems and academic information analysis theory.
The experimental part of the programme is being undertaken by co-workers in the Mechanics of Multi-phase Materials group within the Materials Science Department in collaboration with industrial partners. These partners include Rolls Royce, Inco and Whyman Gordan. Attention is initially being focused on the cold forging of aluminium components which are then annealed to give recrystallised microstructures. Subsequent to annealing, the materials are sectioned and metallographically studied in order to determine their microscopic characteristics (e.g. grain size, aspect ratio and orientation). This data is then used to develop neural network techniques to characterise the relationship between the forging conditions and the microstructure of the material. The neural networks are trained using data from forging processes at a range of temperatures, deformation patterns and times. They then effectively interpolate this data to produce predictions of the final microstructure that would be achieved under a wide range of permutations of the forging variables. Ultimately we would like to be able to invert the problem, so that we can deduce the forging parameters required to yield a desired microstructure.
The first part of this project focused on the application of Gaussian process models (which can be considered as a variant of neural networks) for modelling this problem. Some metallurgical results from this project can be found in these papers:
The second part of this project looked at the development of a recurrent neural network architecture for modelling the hot-forging of materials. Hot-forging is a dynamic process in which dynamic recrystallisation of the material occurs during the forging process. This leads to a marked increase in the complexity of the process, which now exhibits path-dependent behaviour. The dynamic modelling problem and the network architecture is described in the following paper. For further details see the dynet web page.
This project was supported by Inco Alloys Limited, Alcan
International, DERA and the EPSRC.
Return to my homepage.