AI-based image recognition of laser-induced surface structures - An approach for machine learning in ultrashort pulse laser processing
R. Thomas¹*, E. Westphal¹, G. Schnell¹, H. Seitz<sup>1,2

1</sup> Chair of Microfluidics, Faculty of Mechanical Engineering and Marine Technology, University of Rostock, Justus-von-Liebig-Weg 6, 18059 Rostock, Germany,
² Department Life, Light & Matter, University of Rostock, Albert-Einstein-Str. 25, 18059 Rostock, Germany
*robert.thomas@uni-rostock.de


One way to create micro- and nano-scale surface structures on various materials is by using ultrashort pulsed laser processing. Using ultrashort laser pulses, self-organized structures can be created in the laser spot during the machining process and transferred to a larger scale by suitable scanning strategies. The application areas of such structures range from technical applications, surface wetting, and optical applications to chemical and biological applications. The formation of the surface structures depends on the laser parameters (such as laser fluence, polarization, scanning strategy, and the number of pulses), the material properties (such as chemical composition and physical properties), and the ambient conditions.
However, since the underlying formation mechanisms of the structures are very complex and depend on several nonlinear relationships, it is challenging to choose the proper process parameters and scanning strategies for a desired surface modification. This poses problems in producing surface structures using ultrashort laser pulses and represents a barrier to using laser systems.
To solve this problem, we introduce a machine learning (ML) approach. As a first step, we present a successful AI-based image recognition of direct-light microscopy images of different structure classes. The approach using a direct-light microscope allows in-line process control and autonomous structure classification of the system using ML. Furthermore, machine learning can be used in the future to work out optimal process parameters autonomously.
For image recognition, we generated three different structure classes (LIPSS, Crater-structure, Micro-structure) compared to the reference on two steel substrates (X37CrMoV5-1 and X5CrNi18-10). To train a pre-trained model, 250 different images of X37CrMoV5-1 steel were used from each structure class (4x). The classification test of the trained models was completed using 30 other images of each structural class of both steel substrates.
The results of the ML-based classification show that the generated direct-light microscope images could be effectively used to train an ML algorithm, which was subsequently able to automatically classify new, unknown images concerning the surface structure present. A web application was developed to interact with the trained ML algorithm, allowing new microscopic images of various surface structures to be analyzed without specialized knowledge. Therefore, the ML-based classification of laser-induced surface structures is very promising for many in-line applications.