Particle Filters for Real-Time Autonomous Search of Chemical Synthesis Parameters (Reference Maskell LRC122)

Particle filters are a state-of-the-art algorithm for processing incoming data streams. Recent research at Liverpool, funded to support international collaboration between defence agencies in the UK, US, Canada and Australia, has shown that particle filters offer impressive performance in contexts which involve pronounced temporal latencies: other approaches struggle in such contexts.

This project will conduct research to understand the particle filters’ utility in a control setting where the particle filter is not only used to infer the current state of the world, but also used to determine the best way to interact with the system to achieve some objective. The specific context that will focus the research (and for which hardware for testing and demonstration of novel techniques will be available) is flow chemistry. In that setting, the flow of the chemical reagents can be controlled in real-time. The reagents mix and react in the flow reactor (i.e., pipe). It then takes an appreciable (and reagent specific) time for the product to flow to sensors which provide real-time measurements of the product, albeit with pronounced latency relative to the control inputs. A particle filter could, in principle, be used to infer the content of the flow reactor from the known controls and observed sensor data. This would make it possible to transform the way materials are designed and made (e.g., facilitating quicker reactions, fine control of parameters, and, crucially, easy scale-up to industrially-relevant scenarios). Researching such potential step-changes is a key objective for the `Leverhulme Research Centre for Functional Materials Design’ within the University’s Materials Innovation Factory (MIF), which is funding this PhD. The MIF and wider project team have strong industrial links to potential users of flow chemistry (e.g., Unilever, Astrazeneca) and providers of flow chemistry systems (e.g., Waters).

In this setting, the objective is to choose the controls to both explore how the mix of chemical reagents affects the properties of the product but also to exploit this understanding to optimise the mixes of reagents in terms of a certain combination of properties. Such a combination of exploration and exploitation is often encountered in the context of autonomous systems and has been extensively explored in neighbouring contexts (e.g., controlling teams of drones to find specific people). There is a close relationship to techniques (e.g., reinforcement learning and Monte Carlo roll out) that have been developed to tackle (easier) problems such as learning to beat all humans at Go.

This PhD will begin by adapting existing particle filters to consider the flow chemistry problem, embedding those particle filters within a real-time autonomous control system, testing the approach in the context of a (real) flow chemistry reactor within the MIF and evaluating the approach in comparison with a baseline approach (where a human defines the control parameters). The PhD will proceed to advance the constituent components of the system.

Qualifications: 2:1 Degree in Computer Science, Engineering, Statistics or a similar discipline.

Please apply by completing the online postgraduate research application form: https://www.liverpool.ac.uk/study/postgraduate-taught/applying/online/

Please ensure you quote the following reference on your application: Particle Filters for Real-Time Autonomous Search of Chemical Synthesis Parameters (Reference Maskell LRC22).

Please note that this PhD includes an opportunity to gain experience of supporting teaching within the school of Electrical Engineering, Electronics and Computer Science. The expectation is that this would involve up to 3 hours per week during term time.