Motivation and Background
New manufacturing processes typically involve a time-consuming development phase at the laboratory scale and a further ramp-up phase at an industrial scale before they are capable of producing with the required quality, productivity and cost-effectiveness. This is especially the case when new materials and new processes are employed, when the manufacturing process and its sub-processes have a vast number of parameters and high-dimensional state-spaces, and when only poor dynamical process models are available at the outset. From an abstract point of view, a huge and complex optimisation problem must be solved in terms of the process structure, parameter values, control strategies, and process instrumentation (sensors and actuators).
Traditionally, experts accomplish this via extensive experiments based on their knowledge, experience, and intuition. But acquiring experimental data from manufacturing processes is associated with large costs due to, e.g., downtimes of expensive infrastructure, resource consumption for non-marketable output and costs for destructive testing of output products in dedicated laboratories.
Hypothesis: We believe that the systematic use of AI leads to significant improvements in the maturation of manufacturing processes, both in terms of the solution quality and the resource investment needed to arrive at a pursued maturity. However,the high complexity and stochastic nature of real-world manufacturing processes, coupled with a limited number of expensive real-world samples, requires methodological improvements also in the field of artificial intelligence itself.
Scientific and technical approach: First, we strive to improve the state-of-the-art in data-driven model identification, optimisation and reinforcement learning with the goal of minimizing the number of real-world samples required for manufacturing process optimisation. This is achieved (a) by the deep integration of data-driven model identification with prior process knowledge and engineering models, as well as (b) the targeted online design of experiments to create samples with high expected information gain. Secondly, we create a broadly applicable methodology for the AI-assisted transformation of an initially immature manufacturing process towards an applicable and mature one. Ingredients are inter alia a temporary systematic over-instrumentation, a process modularisation, and an iterative maturation procedure. This is a step towards a professionalization of AI in the sense of an engineering discipline with procedures and tools that can scale to large and heterogeneous teams.
Relevance: A much faster transformation of an immature manufacturing process into a mature process would allow for a much shorter time to market for new products, and would therefore provide competitive advantages for companies. Although difficult to predict, we believe that by utilising the intended methodology, a time and cost reduction in process development up to 50% can be achieved.