Prediction of damage evolution in forged aluminium metal matrix composites using a neural network approach

This paper investigates the feasibility of using neural networks (NN) and Gaussian processes (GP) for the modelling of damage evolution during the forging of Al–SiC particle reinforced brake discs. Such models are essentially multi-parameter non-linear interpolation tools and can be ‘trained’ to establish non-linear relationships between the input parameters (here the forging conditions) and microstructural outputs (in our case the level of damage).

Models were trained on the levels of damage found in a brake forging. Traditionally, predictive tools for microstructure evolution have relied heavily on a large database of isothermal, constant strain rate tests representing the range of processing conditions (strain-rate, temperature, etc.) experienced during forging.

Such databases are both time consuming and expensive to generate. In this paper we combine as-forged microstructure measurements with finite element (FEM) predictions of the processing conditions to develop models based on a minimum of mechanical testing. The sensitivity of the network to both choice, and the number, of input parameters is also investigated.

The predictive capability and the extent to which models over-fit training data are discussed. Finally we predict damage in the brake component using a network trained on a completely different forging.