Professor Alison Walker and her research group in the Department of Physics have developed powerful tools that can simulate perovskite materials and how they apply to solar cells. By combining modelling techniques such as Monte Carlo and drift-diffusion modelling with machine learning, they’ve enhanced their simulation power even further.
Solar cells have traditionally been made using silicon-based semiconducting materials. While it’s one of the Earth’s most abundant elements, the silicon used in solar cells must be purified, making it an expensive material to work with. The resulting solar cells are also large and rigid, limiting their uses; for example, the wide, flat panels seen on rooftops.
Perovskite materials are a family of minerals with a unique crystal structure and a range of surprising properties. Not only are the cells cheap to produce, lightweight and flexible when used on their own, the cells can be built into solar panels to provide a significant increase in performance compared to standalone silicon solar cells.
While perovskite cells clearly show potential, one of the challenges facing physicists is the limited stability of the material. Early cells have been shown to degrade rapidly, becoming unusable within minutes or hours, making them unviable commercially.
Simulating the solution
To optimise stability and performance, fast iteration of fabrication and characterisation of properties is essential. However, current processes to analyse underperformance are complex and time-consuming, slowing down fabrication.
Professor Walker and team have been exploring these challenges and showing how combining machine learning and simulation methodologies allows for much faster and more direct characterisation of materials and devices. Together, they’ve developed a set of computer codes to create a virtual model that can pinpoint causes of features seen in the measurements using only a few hours of computation.
The virtual model is continuously updated to reflect the current output of a fabricated laboratory device. This allows the team to understand the materials processes underlying changes in the devices’ output. By accurately and rapidly simulating the function and performance of the lab device, this opens up the possibility of pinpointing the origins of degradation and allows improvements to be made much more quickly in future device iterations.
Read the papers
Their studies have been demonstrated in two recently published papers. In the first, published in APL Machine Learning, the team showed how these techniques can be used to test hypotheses about the physical processes within these devices. In this case, they address outstanding questions on charge transport in perovskites.
In the second paper, published in JPhys Energy, machine learning solves the inverse parameter problem, where a simulation method (here drift-diffusion) is combined with Bayesian parameter estimation. The result is deducing the input parameters for the model from its output, which could be measured characteristics.
The majority of the work for the first paper was done with the aid of a summer bursary for Samuel McCallum, the lead author of both papers. This allowed him to collaborate with the research group of T.T.-Prof. Dr. Pascal Friederich at the Karlsruhe Institute of Technology. Samuel was an MPhys student in the Department of Physics and is now a postgraduate in the Department of Mathematical Sciences with the EPSRC Centre for Doctoral Training in Statistical Applied Mathematics (SAMBa).
The second paper came from an MPhys project by Samuel and Oliver Nicholls, who was also an MPhys student in the Department of Physics. Coauthors of the two papers are Jamie Lerpiniere, a postgraduate student in Professor Alison Walker's group; Matthew Cowley, a postgraduate student in Professor Petra Cameron's group in the Department of Chemistry and Dr Kjeld Jensen who works for BT.