Deep learning for 3D modelling of multiple pulse femtosecond ablation

Mills, Benjamin, McDonnell, Michael, David Tom, Heath, Daniel, Xie, Yunhui, Grant-Jacob, James, MacKay, Benita, Scout, Praeger, Matthew and Eason, Robert (2019) Deep learning for 3D modelling of multiple pulse femtosecond ablation. In 15th International Conference on Laser Ablation (COLA) 2019.

Abstract

Femtosecond laser ablation can enable extremely high precision materials processing, as multiphoton processes can be leveraged to produce single features that are considerably smaller than the laser wavelength [1]. However, due to the extremely nonlinear nature of femtosecond laser machining, results can be extremely difficult to predict, and instead a systematic exploration over all possible laser parameters is generally needed in order to identify the optimal parameters for machining. Femtosecond laser ablation is extremely complex; it incorporates optics with processes such as thermal expansion, phase change, fluid dynamics, ionisation and chemical reactions. Such calculations quickly become too complex to compute at experimentally useful dimensions. What is needed, therefore, is a simulation methodology capable of precisely modelling femtosecond laser machining in a computationally efficient manner.

Neural networks have shot to prominence in the past few years, due to their effectiveness in solving complex problems without the need for human understanding of the underlying physical principles [2]. Rather than being programmed with equations, neural networks can “learn” how to simulate a physical system directly from processing experimental data. We have recently shown that a neural network can learn how to predict the 3D surface profile of a substrate after being laser machined with a single laser pulse, for any laser spatial intensity profile [3]. We subsequently quantitatively demonstrated that the neural network had learnt the properties of optical diffraction [4].

As shown in Fig. 1, we have now extended this work, in order to demonstrate that a neural network can also learn the physical rules that govern the laser machining of a substrate with multiple laser pulses, even in the case where each laser pulse has a different spatial intensity profile. The predicted 3D depth profile shows extremely strong agreement with the experimentally measured 3D depth profile. This specific example was not included in the training of the neural network and hence this accuracy will be typical for any combination of spatial intensity profiles. The neural network also explicitly demonstrates that the order of the spatially shaped pulses influences the outcome of laser machining, where, for example, machining coarse features before fine features maintains high edge quality of the fine features, whereas the opposite does not. We anticipate that the capability for spatial and temporal modelling at such high accuracy will be a key enabling technology for efficient femtosecond laser ablation.

This record has no associated files available for download.

Contributors

Author: Benjamin Mills

Author: Michael, David Tom McDonnell

Author: James Grant-Jacob

Author: Benita, Scout MacKay

Author: Matthew Praeger

Author: Robert Eason

Download statistics