Task Group F2 – Cluster of Excellence PhoenixD

Our motivation

Companies are facing decreasing batch sizes and decreasing product life cycles. This also applies to manufacturers of optical technologies. At the same time, complex multi-step processes often characterize the production of optical components. Highly automated self-optimizing systems are therefore required in the future, in order to be able to manufacture optical components efficiently, even with a large number of variants.

Nowadays, adaptive manufacturing keeps predetermined design parameters within the tolerance range already during production through active control. For the optics production of the future, however, PhoenixD is pursuing a different vision: the change from optimizing static process parameters to optimizing the function of the component in real time during production. A virtual model ("digital twin") of each optical component's individual properties and performance is to be created during the process.

In the future, it should be possible to produce customized optical systems economically in very small quantities.

For this vision to become reality, the members of the task group research systems for the rapid acquisition of the optical properties of manufactured components. The processing of real-time data and individual production history in production control is carried out with machine learning methods. This generates process knowledge, which can also be used offline to adapt the processes. In the future, it should be possible to produce customized optical systems economically in very small quantities.

Our Research

The adaptive planning of production processes and the ability to react to unforeseen changes requires different control loops in machine tools (Figure 1).

Figure 1: Control loops of self-optimizing machine tools

The inner control loop contains systems that support short-term autonomy. These include process monitoring systems and process control systems. A major requirement for the systems is real-time capability. Accordingly, the systems assigned here are assumed to have online capability.

In the outer control loop, on the other hand, no real-time capability is required, since process adjustments are carried out with the component cycle. These control loops support a medium-term autonomy. Elements of the outer control loop often include approaches for data fusion and subsequent storage of process data in databases. The continuously growing database is the basis for process models that can be used for adaptive or self-optimizing process planning.

A polishing process currently serves the task group as a model system for a self-optimizing, adaptive manufacturing process. Figure 2a shows the measurement setup consisting of a photographic lamp and a digital camera that records the work piece. Using the established Cook-Torrance model, the light transport from the source through the test setup to the camera is simulated and the error between the image (Figure 2b) and the simulation is optimized according to the desired parameters, in this case roughness (Figure 2c).

Figure 2: a) Setup of the measurement system, b) Recorded image of a polished surface, c) Simulated image of the surface

Process models are required in order to link the process result with the actuating variables of the process and to identify suitable actuating variables a-priori in the further course. There is a high-dimensional, non-linear functional relationship because a large number of influencing variables determine the complex physical relationships in surface generation.

Methods of machine learning represent a promising approach for the modelling and continuous updating of such complex relationships. In the work of the task group, artificial neural networks (ANN) and decision trees have been analysed.

The approach is to be transferred to other manufacturing processes in current and future work. Thereby, the focus will be on ultra-precision machining with diamond tools as well as the additive manufacturing of glass materials.

In addition to measuring the resulting surface quality, machine and sensor data from the process will be used to improve the model quality. For this purpose, a simulation accompanying the process is to be used.

A further research topic is the development of a virtual workpiece model. It is known that virtual workpiece models are excellently suited for information feedback throughout the product life cycle. Therefore, a virtual workpiece model is to be developed in cooperation with the departments S and M, which allows an estimation of the performance of the individual optical system already during production.

The basis of adaptive planning is comprehensive production data and fast inline measuring procedures, which are transferred into process knowledge via machine learning methods.

The research objectives of the group are closely interlinked with other groups in the Cluster of Excellence. For example, the simulation methods of group S1 will be used in future to simulate the performance of the manufactured components with help of the virtual workpiece model already during production. Innovative measuring systems from group F1 offer new possibilities for characterizing the manufactured optical systems during production. The work of group M2 includes research on a process simulation for additive manufacturing, which will also be part of the control loops shown in future. Furthermore, the group cooperates with group M4 regarding the integration of the research results into machines and into the Manufacturing Grid.