Predicting alloy microstructures using Physics-driven machine learning

Titanium based alloys are used in a range of high performance applications in aerospace, ranging from jet engine components, to airframe or parts of the landing gear.

During the production of wrought titanium components, the alloy can undergo a number of heat treatments and hot working cycles, and optimal properties for high performance applications require precise control over these processing conditions.

However, during these processes, the titanium alloy undergoes several cycles of deformation, recrystallisation and phase transformation which can happen simultaneously, resulting in heterogeneous deformation and unstable microstructure evolution.

The aim of the project is to extend physical based models for microstructure evolution by applying machine learning techniques to microstructure data provided by Lightform project partners.

The problem being address in this project in determining the most effective algorithms for linking modelling of fundamental Materials Science with real world microstructural and thermomechanical datasets. A full physics-based model can be computationally expensive and will still require validation and continuous refinement against real world data. On the other hand relying entirely on experimental methods to determine the effect of processing on microstructure would be extremely limiting, and dismiss the enormous body of knowledge surrounding Metallurgy. Therefore a combined approach applying statistical methods to physics-based modelling should provide the most effective way of predicting the best thermomechanical processing methods.

This is conducted through a systematic review of statistical inference methods, and related algorithms, and their integration with physics-based modelling. The output of this would ideally be a computationally efficient tool for predicting microstructure which could guide the development of new alloys and forming routes.

The work initially looks at crystal plasticity, and using statistical inference on experimental data to find crystal parameters. These can be used physical modelling to predict mechanical properties of single phase metals. The work will then be expanded to use statistical methods to determine parameters that can be used to determine microstructures post thermomechanical processing. Ideally it will be possible to predict grain size, formation of different phases and texture.

This project is part of a larger 'Doing More with Less' whose final outcome will be the development of a predictive model capable of accelerating developments of metallurgical manufacturing methods.