Cognitive Chemical Manufacturing

It is an unceasing challenge to reduce the time scale for development of new chemical products to the point of reliable manufacture and entrance into the market place. Any delays will result in both the loss of revenue for companies, and delayed benefit to the consumers, whilst rushed development might lead to quality issues. This can have significant societal implications, for example impacting the availability, or critical quality of crucial medicines to patients. Therefore, there is a real need to minimise the time taken to identify safe and robust chemical manufacturing processes. These processes however, are complex with process outcome being affected by a vast number of chemical and physical parameters; e.g. temperature, pressure, reagent stoichiometry, pH, heat and mass transfer affect quality and scalability making the definition of a chemical process at manufacturing scale a very challenging task. The sheer number of variables means that a systematic, 'change one factor at a time' approach is practically impossible and generally disregards the fact that some factors might be heavily correlated.

This exciting project combines the expertise of IBM in the development of algorithms for optimisation and the use of automated model generation and discrimination by researchers at UCL with the experimental automation expertise within the Institute of Process Research and Development at Leeds and the use of advanced hydrothermal reactors developed at the University of Nottingham. This research capability will be used to develop new algorithms for machine learning based generation of chemical process design knowledge and coupling these algorithms to a cyber platform for automated experimentation. The combined cyber-physical system will be validated via in-depth case studies related to current manufacturing challenges faced by AstraZeneca, a large UK based manufacturer of Pharmaceuticals who are the UK's fifth largest exporter and Promethean Particles, a SME who have recently opened their first nanoparticle manufacturing facility.

This project aims to develop an Industry 4.0 approach revolutionising the transfer from laboratory to production using advanced data-rich and cognitive computing technologies. We will develop new algorithms based on Bayesian Optimisation and evolving Kinetic Motifs that merge data analysis and the generation of further experiments. Cloud based machine learning services (hubs) will generate experiment setpoints delivered through the cloud to automated laboratory platforms (LabBots). A key novelty is that the analysis services can receive and analyse results, and post further experiments to the LabBots, thus generating a data generation - data analysis closed-loop. This enables the application of machine learning to chemical development: the system will continuously learn, increasing in confidence and knowledge over time, from previous iterations.

Using the same cloud based platform, this process understanding can be rapidly transferred to PilotBots; production scale manufacturing robots that use the same data transfer protocols, but on a 102-105 times larger scale. This fully automated approach has the power to reduce the cost, improve quality and robustness and minimise development time; bringing products to market faster and therefore enabling the beneficial effects to be realised more rapidly.

Our approach will enable the design, selection and evaluation of manufacturing process and technology based on mechanistic and statistical data models. Further, and not less important in development, pilot quantities are easily generated, supporting late stage development activities (e.g. efficacy and stability testing) and the same data analysis services can reconcile the pilot and lab data. The anticipated impact of this approach will be demonstrated on real world manufacturing challenges faced by our pharmaceutical and nanoparticle producing industrial partners.

The development of automation and machine learning for manufacturing has been identified as a critical pathway towards establishing Industry 4.0, 'Smart Factories'. The project aims to enable the design of smart factories by closing the loop between automated experimentation and machine learning. This will bring big data upstream in the development of manufacturing processes allowing a richer choice of both process equipment and process conditions. On the one hand this leads to innovative, more sustainable processes with better quality, less variability, lower hazards, on the other it broadens the range of transformations available to the product designer thus widening the so called 'chemical space' (for instance enabling the "forbidden chemistries" investigated by Ley et al.). Project success means such chemistries can be investigated quicker, earlier in development and in more depth increasing the likelihood of scale-up. The potential economic impact of the project within the UK through manufacture of new successful products is measured in billions.

The development of a new distributed cyber-physical platform, in which the developer interfaces via a cognitive machine learning system with the experimental platform and its operators will enable the equivalent revolution that parallel processing and data centres has brought to computing. Our vision is one of many independent suppliers of experimentation, connected via a thriving international market space to digital service providers of machine learning systems. This forms an exciting new cyber-physical industrial ecosystem. Such an ecosystem will also induce significant demand of the flow equipment industry, in which the UK is very well represented (e.g. Uniqsys, Vapourtec, AM Technology and ChemTrix, Cambridge Reactor Design), and may open up a new market for cyber enabled hardware systems like the LabBot and PilotBot.

To achieve the above impacts requires a steady supply of highly trained professionals who apart from their chemical manufacturing domain knowledge have significant skills in the digital domain. These are in short supply (e.g. see AZ letter of support) and the project aims to train 7 people in this area, and will expose many more students (undergraduates via MSc, MEng and MChem projects, PhD via peer learning etc.). It is expected that most of these scientists will gain employment in manufacturing companies, thus, introducing them to the concept we have developed.

Impact is also generated through the projects industrial partners: Swagelock are keen to be involved in commercialisation of the physical part of the system; AZ is planning to embed the concepts developed in a new software to be commercially developed with an aim of wide availability the long term; PP aims to innovate commercial manufacturing processes and IBM plans to use its established program for the development of Industry 4.0 (Watson IoT) to commercialise its research.

Society at large will benefit from innovated, personalised, safer and more sustainable end-user products, as well as additional R&D and manufacturing jobs. The academic impact of the novel BayesOpt algorithm developed by IBM/Hartree and the UCL/UoL kinetic motif algorithms and the coupling of these algorithms in a closed loop to automated experimentation will be fundamental in the analysis and generation of large datasets. This capability will cross over into other areas like advanced and nano material design, on demand medicines and end product formulation. Linking equipment to a market place allows more users access and thus increases equipment utilisation. In academia, this will reduce the cost of research to the taxpayer and in industry it leads to lower cost development.