Development of an Artificial Intelligence for the control of an Atomic Force Microscope has been started

The technique of artificial intelligence (AI) has developed rapidly over the last two decades. This has led to AI-based systems becoming very powerful in performing a wide range of tasks. However, unlike most classical approaches, whose tasks are now increasingly being performed by AI systems, AI algorithms do not behave deterministically or predictably. This poses a considerable challenge, especially for the application of AI in metrology, where the reliability and understandability of results are of particular importance.

With these aspects in mind, the development of an AI for controlling the distance between the probe tip and the measurement surface on an atomic force microscope (AFM) was started. The aim of the development is to extend or replace the PID control used up to now with an AI, adapted to the dynamic behaviour of the AFM system. The expected advantages of the application of an AI-based control are significantly reduced control deviations compared to an (optimised) PID control and, above all, faster reaction characteristics with simultaneous avoidance of overshoots.

The development of the AI was started on a computer model of the AFM. The expected advantages from using an AI as a controller could already be demonstrated. Since the AI is to be trained later on the real AFM itself for optimal performance, a strategy must be developed to ensure that no critical control states occur during training, as the AFM tip and the sample could be damaged because of them colliding harshly. The approach currently being tested is a hybrid control of AI and classical PID control. In this approach, the AI is allowed to take over the control and optimise its behaviour until a certain (not yet) critical control state is reached. Once this limit is exceeded, the PID control takes over until the AFM is again in a well-controlled state. This hybrid approach is fundamental for the training phase, in which the AI is not yet a functioning controller, but also for continuous operation this approach enables the combination of the performance of modern AI with the reliability of classical control approaches.

In order to be able to assess the reliability of the AI itself, the so called "deep Q learning" technique is used. With this technique, the feedback signal is not simply calculated, as would be expected for a control system, but the AI estimates the "quality" for a discrete number of possible feedback values for a given control state and selects the feedback value with the highest "quality". If we plot the estimated qualities of the possible feedback values over the values of the feedback signal, we obtain a “quality curve” that enables the assessment of the AI's understanding of the system behaviour. The goal is to obtain a smooth "quality curve" that indicates that the AI has learned even small differences in the effect of similar feedback values on the control state. If the AI estimates smooth "quality curves" for a wide range of control states, it has reliably learned the system behaviour and is accordingly suitable for application.

Figure 1: Example of a "quality curve" for possible feedback values (here, voltage changes for the piezo stage of the atomic force microscope) for a given control state. In the example shown, the quality curve is not yet smooth, so AI would not yet be considered for practical application.