Todays scanning probe microscopy techniques already allow us to poke, prod, push, probe and interact at the atomic level. However, this involves spending hours upon tedious hours of busywork trying to gradually (and semi-randomly) coerce our probes into a vaguely usable form.
Not only does this raise very good concerns about the effect of human bias in analysis, but it massively reduces our ability to experiment reliably. In effect, we build an increasingly unstable house ofcards, then painstakingly rebuild it over and over until we get what we decide is good enough. The limiting factor is not one of precision, but of patience, luck, reliability, and trust.
However, where human stubbornness and conventional computing techniques have failed, machine learning concepts with roots dating back to WW2 can become a key method of asserting our control over the nanoscale. Key to this are neural networks, which, thanks to significant increases in computing power and widespread adoption of free, open source tools such as Python and Tensoflow, have resulted in fundamental improvements toa huge variety of predictive, classication and strategical tasks.
Research Areas
Towards Automated Probe Microscopy
Ultrahigh vacuum (UHV) scanning probe microscopy enables arguably the ultimate control of condensed matter (via single atom manipulation/positioning) and has been associated with some of the most exciting, inspirational, and elegant experimental work in condensed matter science. Unfortunately, it can also be one of the more frustrating techniques used in experimental science, with many fruitless hours spent on attempting to coerce the probe into a structure that yields, retains, and accurately reproduces atomic resolution images. Collectively, vast numbers of researcher person-months in the scanning probe and nanoscience research communities are therefore currently wasted in conditioning STM and AFM tips (during the scanning process) via voltage pulsing, field emission, high bandwidth scanning, high (or low) feedback loop gains, and other "tricks of the trade".
We are developing, through a joint Physics - Computer Science collaborative approach, a "one click" set of algorithms and protocols which will instead automate the optimisation of tip structure (and tip chemistry), representing a radical, and dramatically more efficient, departure from conventional working methods in scanning probe experimentation. This video shows the automatic optimisation of an STM tip to provide atomic resolution of the Si(111)-(7x7) surface, involving no human operator involvement (other than moving the tip to the sample surface).
Since then, we have been working to use convolutional neural networks, and a combination of supervised and unsupervised machine learning to not only automatically assess the specific state of an STM tip in real time, but eventually, and intelligently, coerce the tip into a desired state.
Regardless, once we acquire our data we still have the issue of subjectively processing all ofit; unlike computers, humans cannot perform billions of calculations per second, and worse still are inherently biased, bottlenecking the research process. Machine learning again provides an exciting means to make better use of the abundance of data SPM creates.
Indeed, while one of the greatest advantages of atomic force microscopy (AFM) is its ability to produce huge numbers of scans on avariety of distinct surfaces, the need to manually select data to analyse leads to much of it being under-used at best. We are only beginning to find ways of better using existing data left unanalysed onold hard drives, CDs and even oppy discs. Further, the requirement to manually pre-process data raises uncomfortable questions; at what point could our desire for aesthetic quality of publishable figures cause us to introduce unphysical information?
As such, we have demonstrated that it is possible to use simulated AFM images as automatically labelled training data for a neural network, which then correctly generalises to real AFM scans. We employ an automated,optimised pre-processing routine for real images of specific structures, combining this with a denoising autoencoder to provide effective, automated binarisation of realimages. We then combine these systems to quickly find experimental AFM images sowing specific structure growths in old repositories of data.
When performing SPM techniques on surfaces, we often take multiple spectra to better understand the surface we are looking at. However, tip-apex flaws (especially in the case of poorly understood spectra) can cause confusing, inconsistent and non-linear artefacts, and distinguishing between what information is physically real, what is artefact, and what is noise, is often extremely difficult. This is especially true when considering that we often generate large amounts of spectra.
As such, we aim to use unsupervised clustering techniques such as k-means, in order to automatically distinguish between STM strectra of different molecules and/or tip states.