In the “AI-Driven Smart Factory” as propagated by SEMI in the Smart Manufacturing Roadmap, Artificial Intelligence, and in particular Reinforce Learning techniques are a distinguishing enabling element along the transition from today’s Smart/Industry 4.0 approaches to the future “Smart 2.0 Factory” and especially the “Autonomous Control Room”.
This raises the question for what kind of tasks and functions such techniques can add value on top of what is possible with established methods, which in this presentation will be discussed from the Industrial Engineering point of view, i.e. decisions around capacity planning and material flow optimisation in a semiconductor wafer fab.
Whenever such a decision is to be taken, the expected performance of the selected option with regard to a particular objective against alternative options is required. In the case of a wafer fab, the underlying objective function should be a capacity model that is able to represent a Discrete Event Logistics System, and since in most cases the interdependency between capacity and cycle time, i.e. the causality between a certain solution option and its impact on capacity and cycle time needs to be considered, this capacity model should be (in the case of a schedule to be generated) a deterministic or (in the case of a longer term plan to be generated) a stochastic Discrete Event Simulation model, also because these causalities need to be portrayed with sufficient fidelity.
To determine an optimal plan or schedule, powerful optimisation techniques are needed, requiring detection of correlations between decision variable values and objective values. That is where Reinforcement Learning techniques come into the picture.
The presentation will also explain why such Reinforcement Learning techniques can be more useful for scheduling but may have limited value for situation-based dispatching. Also, in an environment where the underlying capacity model is always a simplified representation of actual operations, an optimal solution would never be found anyway. Rather, for all practical purposes it is sufficient to determine a much better solution with as few iterations as possible. The presentation will also illustrate how this is enabled through the solution approach of the D-SIMCON Digital Twin framework.
For further details please click here.
For further enquiries please contact:
Dr Peter Lendermann, Chief Business Development Officer, peter@d-simlab.com