Predictive visualisation of laser processing via deep learning

Abstract

Laser materials processing is a non-contact manufacturing technique that has applications in many areas of industry and academia. However, due to the highly nonlinear nature, analytical and numerical modelling of laser processing can be extremely challenging. In recent years, deep learning has rapidly become a critically important tool for data-driven modelling of complex systems. This thesis therefore seeks to apply the latest advancements in deep learning for the purpose of predicting, modelling and analysing a range of laser materials processing phenomena.
In this thesis, deep learning techniques were used to predict the appearance and topography of materials cut with continuous wave and pulsed laser sources. The neural networks were able to correctly predict the appearance and topography of the edge of laser cut metal samples, as well as the depth profile of ablated holes produced via femtosecond machining. These techniques enable the prediction of laser cutting defects and ablated hole depth profiles, as well as enabling the simulation of large ranges of configurations through parameter interpolation, all of which lead the way towards automated experimental optimisation. Neural networks were also shown capable of unsupervised identification of underlying experimental parameters, a result which offers an early demonstration of the potential for deep learning in data-driven scientific discovery.

More information

Identifiers

ORCID for Alexander Courtier: orcid.org/0000-0003-1943-4055

ORCID for Michalis Zervas: orcid.org/0000-0002-0651-4059

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

Catalogue record

Export record