Selective laser cleaning of microbeads using deep learning

Laser cleaning is widely used industrially to remove surface contaminants with high precision. Conventional methods, however, lack real-time monitoring and feedback loops, often necessitating over-machining to ensure complete contaminant removal, which leads to inefficient energy use and potential substrate damage. In this work, we demonstrate a concept of selective laser cleaning via the application of femtosecond laser pulses and polystyrene microbeads with a diameter of 15 μm. These microbeads model challenging scenarios in high-precision optical work and delicate surface treatments across laboratory and production settings. To enable adaptive, real-time cleaning, we integrated a neural network that predicts the sample's appearance after each laser pulse into a feedback loop, tailoring the cleaning process to a bespoke target pattern. This method ensures precise contaminant removal with minimal energy use, making it highly promising for applications demanding strict material control, such as wafer cleaning, sensitive surface treatments, and heritage restoration. By combining machine learning with ultrafast laser technology, our approach significantly enhances the efficiency and precision of cleaning processes.