Microstructural fingerprint: The application of machine learning methods for the characterization and optimisation of electrode microstructures

点击翻译按钮获取中文摘要

Applications are invited for a research studentship in the field of machine learning for energy storage applications leading to the award of a PhD degree. The post is supported by a bursary and fees (at the UK/EU student rate) provided by the The Faraday Institution. TFI Cluster PhD students receive an enhanced stipend over and above the standard EPSRC offer. The total annual stipend is approximately £20,000 (plus London weighting) plus an additional £7,000 annually to cover training and travel costs. Recipients will have access to multiple networking opportunities, industry visits, mentorship, internships, as well as quality experiences that will further develop knowledge, skills, and aspirations. EPSRC candidates should fulfil the eligibility criteria for the award. Please check your suitability at the following web site: http://www.epsrc.ac.uk/skills/students/help/Pages/eligibility.aspxThe performance of lithium ion batteries is linked to the 3D microstructure of their porous electrodes. Advances in the field of micro/nano-tomography have enabled researchers to capture the morphologies of these microstructure at a resolution relevant to needs of multiphysics simulation [1]; however, the robust characterisation and analysis of this data remains a challenge. Recent advances in machine learning have seen the development of novel image generation tools. In particular, these include style transfer using hierarchical neural architectures [2], variational autoencoders [3] and adversarial methods [4]. These concepts have been developing rapidly in the context of 2D colour images over the past 5 years but have rarely been applied to the generation of 3D labelled microstructural data. This project would seek to transfer the power of these methods to the field of microstructural analysis and generation. First by enabling the extraction of a compressed representations (a "fingerprint") of these memory intensive 3D tomography volumes and then using these representations to more efficiently explore the space of possible microstructure to find new optimal configurations. This will link up with the significant tomographic investigations underway in both the multiscale modelling and degradation fast-start projects, as well as interacting with the continuum modelling efforts seeking to build simplified models explaining cell performance.