The University of Southampton
University of Southampton Institutional Repository

Thinking lasers

Thinking lasers
Thinking lasers
Advances in lasers now allow the laser-based processing of almost any material, and consequently innovation in this field is becoming heavily focussed on making existing processing techniques more precise and efficient.

Deep Learning (DL), a computing paradigm inspired by biological neurons that learns directly from real-world data, has seen a dramatic rise in interest over recent years, due to its capability in solving extremely complex problems.

The question is, how can these two fields be combined?

The team at the Optoelectronics Research Centre (ORC) at the University of Southampton are addressing this specific question, and exploring whether the power of DL can be harnessed in order to improve the repeatability, precision, and control of high-precision femtosecond laser machining of objects on the micro- and even nanoscale.

The objective of the research at the ORC is two-fold.

Firstly, to explore the application of DL for recognising, in real-time, unexpected events (e.g. laser power fluctuations or unexpected debris), whilst also providing process control (e.g. stopping a manufacturing process at exactly the optimal time, even though the process time may be unknown).

Secondly, to investigate the application of DL for predictive capabilities for laser machining, in order to be able to accurately predict what a sample would look like after machining with any combination of laser machining parameters.

The intention is that these two approaches will be combined, as unexpected events must first be observed before a predictive capability can determine which parameters will need to be changed in order to compensate for the earlier manufacturing error. The critical point is that this combined capability needs to operate in real-time.

DL, which generally refers to the application of neural networks (NNs), is based on the premise of computers learning how to solve a problem by themselves, and therefore there is no need for the inclusion of any equations that describe, for example, the interaction of light and matter, or the probabilistic nature of debris production. This is a significant advantage, as the interaction of laser light with materials, particularly for femtosecond pulses, is extremely complex. Instead, the DL approach simply involves the collection of experimental data, e.g. images of laser machined samples for a wide variety of different experimental parameters. The NN is then trained directly and automatically from this experimental data. Once trained, the NN can process input data and provide useful information on a time scale of tens of milliseconds, and hence is applicable for real-time data processing.

A NN can therefore be considered as a transfer function, which converts input data to output data. In the case of laser machining, the input data could be, for example, spectral data, camera images, and temperature measurements, whilst the output could be anything from enabling beam power corrections, to 3D predictions of the sample surface after machining.

Whilst NNs have been studied and understood for many decades, the rate of adoption across academia has been extremely rapid since 2017, due to the available computing power and data (which both increase exponentially) reaching a critical level. As such, most previous work in this area whilst academically interesting, offered limited potential for industrial application. However, as shown here, the computing power available now allows for accurate and real-time capability, and is attracting significant industrial interest.

Monitoring and Process Control



The schematic describing our setup for the monitoring and process control via a NN is shown in figure 1. Femtosecond laser pulses were focussed down to a spot size of approximately 30 microns, whilst the sample was continuously imaged by the camera. The camera images were then processed by the NN, which could output a wide range of experimental information, and also control the laser.

To date, we are able to determine, via a NN, changes in material, fluctuations in laser fluence, and count the number of pulses used for the machining, along with values quantifying the changes in the shape and position of the laser beam on the sample. The accuracy is superior to human eyes, and much faster, taking approximately 10ms to process each camera image.

The NN was also able to control the laser, in order to stop machining exactly at the point of machining through a random (and hence unknown) thickness of copper, whilst not damaging the underlying layer of glass. Such process control could be applied to the laser cleaning of rust, for example, as the thickness of rust at each position would be unknown, whilst zero laser damage should occur to the underlying surface. The NN was not only able to cease laser machining at exactly the point of complete removal of the copper, but was able to predict the time remaining until task completion. Whilst direct interrogation of the internal workings of a NN is extremely challenging, we suspect that the NN was able to correlate the amount of debris and the appearance of the machined surface with some measure of the remaining depth of copper left to machine.

Predictive Capability

Finding the optimal combination of laser machining parameters, such as laser power, laser wavelength, beam size and machining time, for a customer design specification, can be a costly and time-consuming process. Typically, a technician will systemically explore all combinations of laser machining parameters. It would undoubtedly be convenient if a NN could determine the optimal parameters.

Such a NN would need to comprehend everything from the physical equations describing the interaction of light and matter, heat transfer, and the laws of diffraction, through to the properties of the sample and the distribution and probability of debris and burr, through to the particular nuances of the laser itself. However, this complexity actually isn’t a problem, as the NN demonstrated here was able to learn all this, by itself, exclusively from images of laser machined samples.



Although we have only included laser beam shape in a NN, our early results have been staggering. Figure 2 shows a concept of our work, showing a NN transforming a beam shape into a predicted 3D surface map of the laser machined sample. Figure 3 shows the experimental laser beam intensity profile (shaped via a spatial light modulator), along with the predicted and experimentally measured 3D depth profile. It is important to realise that the NN had never encountered anything like this particular shape during the training process, and therefore the accuracy of the predicted 3D surface demonstrated here would apply to other beam shapes. The NN is so effective that it is almost impossible to tell which image is the genuine experimental result. Of particular interest are the slightly raised surfaces in the middle of the laser machined regions. The sample material here, nickel, is known to melt and reform when irradiated with femtosecond laser pulses, and hence the NN had also learnt rules equivalent to fluid dynamics.



The Future

Computing power increases exponentially, as we use the technology of today to build the technology of tomorrow, and it is now the graphics processing unit (GPU) paradigm that is driving the innovation in DL. As of 2018, NNs have demonstrated basic creativity (e.g. see AlphaZero and DeepMind), and hence the potential for NNs to discover new approaches for laser machining is probably not that far away (in fact we are already working on this). By devoting an entire network to a specialised task, NNs have already surpassed many specific human capabilities. Comprehending femtosecond laser machining can now also be added to this list. The point where computers can outperform humans is almost certainly much closer than most people realise.
Mills, Benjamin
05f1886e-96ef-420f-b856-4115f4ab36d0
Mills, Benjamin
05f1886e-96ef-420f-b856-4115f4ab36d0

Mills, Benjamin (2019) Thinking lasers. Laser Systems Europe, (42).

Record type: Editorial

Abstract

Advances in lasers now allow the laser-based processing of almost any material, and consequently innovation in this field is becoming heavily focussed on making existing processing techniques more precise and efficient.

Deep Learning (DL), a computing paradigm inspired by biological neurons that learns directly from real-world data, has seen a dramatic rise in interest over recent years, due to its capability in solving extremely complex problems.

The question is, how can these two fields be combined?

The team at the Optoelectronics Research Centre (ORC) at the University of Southampton are addressing this specific question, and exploring whether the power of DL can be harnessed in order to improve the repeatability, precision, and control of high-precision femtosecond laser machining of objects on the micro- and even nanoscale.

The objective of the research at the ORC is two-fold.

Firstly, to explore the application of DL for recognising, in real-time, unexpected events (e.g. laser power fluctuations or unexpected debris), whilst also providing process control (e.g. stopping a manufacturing process at exactly the optimal time, even though the process time may be unknown).

Secondly, to investigate the application of DL for predictive capabilities for laser machining, in order to be able to accurately predict what a sample would look like after machining with any combination of laser machining parameters.

The intention is that these two approaches will be combined, as unexpected events must first be observed before a predictive capability can determine which parameters will need to be changed in order to compensate for the earlier manufacturing error. The critical point is that this combined capability needs to operate in real-time.

DL, which generally refers to the application of neural networks (NNs), is based on the premise of computers learning how to solve a problem by themselves, and therefore there is no need for the inclusion of any equations that describe, for example, the interaction of light and matter, or the probabilistic nature of debris production. This is a significant advantage, as the interaction of laser light with materials, particularly for femtosecond pulses, is extremely complex. Instead, the DL approach simply involves the collection of experimental data, e.g. images of laser machined samples for a wide variety of different experimental parameters. The NN is then trained directly and automatically from this experimental data. Once trained, the NN can process input data and provide useful information on a time scale of tens of milliseconds, and hence is applicable for real-time data processing.

A NN can therefore be considered as a transfer function, which converts input data to output data. In the case of laser machining, the input data could be, for example, spectral data, camera images, and temperature measurements, whilst the output could be anything from enabling beam power corrections, to 3D predictions of the sample surface after machining.

Whilst NNs have been studied and understood for many decades, the rate of adoption across academia has been extremely rapid since 2017, due to the available computing power and data (which both increase exponentially) reaching a critical level. As such, most previous work in this area whilst academically interesting, offered limited potential for industrial application. However, as shown here, the computing power available now allows for accurate and real-time capability, and is attracting significant industrial interest.

Monitoring and Process Control



The schematic describing our setup for the monitoring and process control via a NN is shown in figure 1. Femtosecond laser pulses were focussed down to a spot size of approximately 30 microns, whilst the sample was continuously imaged by the camera. The camera images were then processed by the NN, which could output a wide range of experimental information, and also control the laser.

To date, we are able to determine, via a NN, changes in material, fluctuations in laser fluence, and count the number of pulses used for the machining, along with values quantifying the changes in the shape and position of the laser beam on the sample. The accuracy is superior to human eyes, and much faster, taking approximately 10ms to process each camera image.

The NN was also able to control the laser, in order to stop machining exactly at the point of machining through a random (and hence unknown) thickness of copper, whilst not damaging the underlying layer of glass. Such process control could be applied to the laser cleaning of rust, for example, as the thickness of rust at each position would be unknown, whilst zero laser damage should occur to the underlying surface. The NN was not only able to cease laser machining at exactly the point of complete removal of the copper, but was able to predict the time remaining until task completion. Whilst direct interrogation of the internal workings of a NN is extremely challenging, we suspect that the NN was able to correlate the amount of debris and the appearance of the machined surface with some measure of the remaining depth of copper left to machine.

Predictive Capability

Finding the optimal combination of laser machining parameters, such as laser power, laser wavelength, beam size and machining time, for a customer design specification, can be a costly and time-consuming process. Typically, a technician will systemically explore all combinations of laser machining parameters. It would undoubtedly be convenient if a NN could determine the optimal parameters.

Such a NN would need to comprehend everything from the physical equations describing the interaction of light and matter, heat transfer, and the laws of diffraction, through to the properties of the sample and the distribution and probability of debris and burr, through to the particular nuances of the laser itself. However, this complexity actually isn’t a problem, as the NN demonstrated here was able to learn all this, by itself, exclusively from images of laser machined samples.



Although we have only included laser beam shape in a NN, our early results have been staggering. Figure 2 shows a concept of our work, showing a NN transforming a beam shape into a predicted 3D surface map of the laser machined sample. Figure 3 shows the experimental laser beam intensity profile (shaped via a spatial light modulator), along with the predicted and experimentally measured 3D depth profile. It is important to realise that the NN had never encountered anything like this particular shape during the training process, and therefore the accuracy of the predicted 3D surface demonstrated here would apply to other beam shapes. The NN is so effective that it is almost impossible to tell which image is the genuine experimental result. Of particular interest are the slightly raised surfaces in the middle of the laser machined regions. The sample material here, nickel, is known to melt and reform when irradiated with femtosecond laser pulses, and hence the NN had also learnt rules equivalent to fluid dynamics.



The Future

Computing power increases exponentially, as we use the technology of today to build the technology of tomorrow, and it is now the graphics processing unit (GPU) paradigm that is driving the innovation in DL. As of 2018, NNs have demonstrated basic creativity (e.g. see AlphaZero and DeepMind), and hence the potential for NNs to discover new approaches for laser machining is probably not that far away (in fact we are already working on this). By devoting an entire network to a specialised task, NNs have already surpassed many specific human capabilities. Comprehending femtosecond laser machining can now also be added to this list. The point where computers can outperform humans is almost certainly much closer than most people realise.

Full text not available from this repository.

More information

Published date: 20 March 2019

Identifiers

Local EPrints ID: 429553
URI: https://eprints.soton.ac.uk/id/eprint/429553
PURE UUID: 877d8a96-b185-4699-b6ec-1b7abac53aa9
ORCID for Benjamin Mills: ORCID iD orcid.org/0000-0002-1784-1012

Catalogue record

Date deposited: 29 Mar 2019 17:30
Last modified: 30 Mar 2019 01:34

Export record

Download statistics

Downloads from ePrints over the past year. Other digital versions may also be available to download e.g. from the publisher's website.

View more statistics

Atom RSS 1.0 RSS 2.0

Contact ePrints Soton: eprints@soton.ac.uk

ePrints Soton supports OAI 2.0 with a base URL of https://eprints.soton.ac.uk/cgi/oai2

This repository has been built using EPrints software, developed at the University of Southampton, but available to everyone to use.

We use cookies to ensure that we give you the best experience on our website. If you continue without changing your settings, we will assume that you are happy to receive cookies on the University of Southampton website.

×