Current projects

Understanding the impact of manufacturing parameters on composite energy materials using machine learning based data fusion of complementary imaging

Lead researchers: Dr Samuel Copper
PI’s: Dr Samuel Copper

Hydrogen electrolysers, fuel cells, batteries, and solar panels all typically contain functional composite materials. The microstructure of these composites significantly influences the performance of the devices they constitute. The microstructures are the result of multistage manufacturing processes in which many parameters need to be specified, such as coating speeds, sintering temperatures, solid mass fractions, and rolling pressures. The relationship between these manufacturing parameters and the resulting microstructures is often highly complex. Although some efforts have been made to simulate these manufacturing processes directly, this may be more challenging in some cases than simulating the performance of the whole device (which is already very challenging!).

As an alternative to simulating manufacturing, new data-driven approaches are emerging that can directly relate microstructures to their manufacturing parameters. For example, the recently developed machine learning concept of conditional generative adversarial networks (cGANs) is gaining increasing attention for their ability to implicitly learn the mapping between high-dimensional spaces. However, these techniques require high-quality training data.

Training data in this scenario would be microstructural image data. This data would need to be high enough resolution to see the smallest relevant features; large enough to contain a representative volume; and of sufficient contrast to accurately capture the distribution of the material phases.

Meeting these three requirements sufficiently well is rarely possible with a single imaging technique. For example, X-ray tomography can capture large volumes, but can be challenged by resolution and phase identification. Alternatively, electron microscopy can achieve much better resolution, but representativity can be a challenge and serial sectioning must be used to capture 3D volumes, which can lead to its own challenges in phase identification in porous materials.

During a Royce Materials 4.0 project in early 2022, my team developed a machine learning based data fusion technique that can take data from complementary imaging techniques (e.g. X-ray and electron) and combine them into a single ideal microstructural dataset. The method brought together machine learning based super-resolution, style-transfer and dimensionality-expansion to generate high-resolution, representative volumes with excellent phase identification. This work has since been accepted for publication in the high-impact journal Advanced Energy Materials (IF: 29)

In parallel to this work, my team also developed and published a new sample preparation route that enables improved phase identification in porous materials during electron microscopy, by in situ platinum infiltration of the pores. This “kintsugi” method has proven to be very effective and Thermo Fisher UK has expressed the wish to build it into their workflow on the latest generation of their Plasma Focused Ion Beam imaging machines (see attached letter). 

In this project, I aim to bring together our machine learning based data fusion methods with the new imaging techniques to investigate the impact of manufacturing parameters on the resulting microstructure of various energy materials. This microstructural data can then be compared to device-level performance to develop a better understanding of their relationships, thus facilitating device-level design and model validation. Furthermore, the microstructural data can also be used to train conditional generative adversarial networks such that the constrained space of manufacturable composites can be rapidly explored in silico.

Bringing together these two novel methods (data fusion and kintsugi imaging) will allow porous composites to be imaged with sufficient quality to observe the full impact of manufacturing parameters for the first time. This is hugely important to the development of the next generation of energy storage and conversion devices. There is a clear pressing need for the optimisation of energy devices and the machine learning and imaging technique necessary to complete this project have both only recently become available, highlighting the timeliness of this project.

  • ThermoFisher
  1. tbc