Image-based neural networks for monitoring and controlling light-matter interactions

Abstract

Recent advances in the field of deep learning have unlocked an abundance of exciting novel scientific applications. The convolutional neural network (CNN), which is a type of neural network optimised for processing two-dimensional data such as images, has been at the forefront of many of these developments. A fundamental capability of the CNN is the ability to extract numerical descriptions directly from real-world observations, which, as shown in this thesis, enables a wide range of techniques based on the real-time measurement, optimisation, and process control of experiments. Critically, the CNN is trained on data (experimental or theoretical), and hence can be applied in cases where the complexity of the experiment may prevent modelling approaches based on a fundamental understanding of the system. This thesis provides four experimental applications of CNNs across the fields of laser machining, laser-based sensing, and laser-based manipulation, for both supervised learning and reinforcement learning. Firstly, that a CNN can accurately identify the type and concentration of microparticles in a solution, directly from processing the scattering pattern when the solution is illuminated by laser light. Secondly, that a CNN can identify the translation and rotation of a laser beam profile during laser machining, and also automatically cease laser machining when tasked with machining through the top layer of a multilayer sample of unknown thickness. Thirdly, that a CNN, via reinforcement learning, can learn how to translate a microparticle through a maze, whilst avoiding collisions with other microparticles, through the use of the optical tweezers effect. Finally, that a CNN, via reinforcement learning, can optimise the toolpath generation for laser machining, when tasked with manufacturing a specific target pattern, whilst also correcting for errors in real-time.

More information

Identifiers

ORCID for Benjamin Mills: orcid.org/0000-0002-1784-1012

Catalogue record

Export record