There are four different tasks that are being researched as part of WP1
Task 1.1 Compartmentalise ML-captured behaviours
Inspired by traditional sub-space system identification methods or eigensystem realisation algorithms, we will capture interlinked physics relevant behaviours via ML models of appropriate complexity. Input/output derived data-driven models are often not causal and parsimonious. Our approach will let us understand better causal behaviours, by understanding the ML behaviour not at a system level but at a physics-relevant level (informed by developments in WP2). This will enable more transparent understanding of the system behaviours and will enable information propagation and extraction at appropriate level for the application of advanced control theory.
Task 1.2 Control system design/synthesis
Will proceed from analysis of the identified/learnt subsystem dynamics. Individual controllers will be designed for each subsystem in order to meet clearly specified performance requirements. Each controller will be developed, and tested in isolation, before applying each one altogether to yield the overall control system. Information gathered from by T1.1 and 1.2 will be used to inform and direct sensor (particularly proxy sensor) development and deployment within WP3,4
Task 1.3 ML information use in input-weighted control
Having a framework for estimating the importance of observable and unobservable physics in the system will also allow the development of advanced control methods with control actions weighted according to the importance, sensitivity and uncertainty of the controllable parameters. As ML progresses, we will inform WP3,4 as to optimum sensor information requirements (i.e. data rates / sensor locations)
Task 1.4 ML-assisted end-to-end whole process multivariate control & resilience framework
Finally, by integrating the ML models with the developed control scheme in real-time, we will create an end-to-end resilience framework that is capable of efficiently rejecting process disturbances caused by unpredictable variance in materials and formulations. This will be implemented within the full manufacturing setting within WP5.