* I have unashamedly stolen this title from my friend and erstwhile colleague at Nottingham, Richard Woolley. Rich, I hope you’ll forgive the plagiarism.
Scanning probe microscopy is my first love when it comes to experimental science. Although I’ve spent quite a bit of time at synchrotrons over the years, I still turn to SPM first and foremost when it comes to probing the structure of matter. After all, what other technique allows us to not only see single atoms and molecules, but to interrogate them mercilessly, reaching down to the level of individual chemical bonds, and pick, prod, push, and/or pull them around a surface? What other technique enables us to capture not only the electronic structure (both filled and empty states in a single “shot”), but the vibrational “wobbles”, the potential energy landscape (of various forms), the probability density, and even the magnetic signature of a single atom or molecule (in parallel, and with energy resolutions that are in essence only thermally limited)?
And, as I suspect my fellow probe microscopists would heartily agree, what other technique is quite so damn irritating, fist-clenchingly frustrating, and hair-pullingly maddening at times?
Probe microscopists can spend hours, days, or sometimes even weeks trying to cajole the component at the very core of the microscope — the tip (or, more precisely, the atomistic structure at the very apex of the tip) — into behaving itself. We use a variety of recipes to modify the tip apex, ranging from rather gentle and delicate indentations (pushing the probe a few angstroms into the surface), picking up a molecule, or rather “tame” voltage pulses of a few volts… to plunging in, burying the probe, and ploughing a furrow across the substrate. And then we scan and hope to see atomic resolution. But even if we see atoms, it mightn’t be the right type of atomic resolution [1]. We might have a double tip (i.e. two atoms are involved in forming the image), or a triple tip, or, something else entirely…
Those images above are all of the atomic structure of the Si(100) surface (as described in this video, and as sketched in the top right corner of the slide, where we’re looking down on the surface. Large circles represent atoms that are “buckled” out of the surface, creating a zig-zag pattern — technically, a c(4×2) reconstruction — of non-planar pairs of atoms (dimers).) In each case, the surface has the same atomic structure — the variations from image to image are purely tip-related. And if we find the tip in a state that doesn’t give us the image we expect or need [2] then we crash, or hammer, or pulse, or bash, and swear repeatedly until we get what we want. Sometimes — usually around five minutes to lunchtime, or 5 pm on a Friday evening — the tip gets better. Other times it gets much, much worse. So we sit there for hours, trying to recover the tip. Or we change it for a new one. And then try to coerce that into showing us atoms, or more than atoms.
But we don’t need to suffer like this. There is a better way. And finding that lower — and hopefully minimum — frustration pathway to better probe microscopy was the subject of a meeting at the Institute of Physics on Friday. Organised by Martin Castell (Oxford) and myself, the theme of the meeting was machine learning for atomic resolution scanning tunnelling mcirscopy (STM) and atomic force microscopy (AFM.) In attendance were scientists from across the UK who each wanted to move our field forward so as to take the pain away (and, of course, consequently do rather more interesting experiments/theoretical calculations as well.) Oxford, Nottingham, King’s College London, Newcastle, Loughborough, University College London, Warwick, Liverpool Physics, St. Andrews, and Swansea were each represented on the day, with apologies from SPM researchers at Bath, Lancaster, Leeds, Liverpool Chemistry, and Cambridge (who were still keen to be involved but were unfortunately otherwise engaged.)
One striking statistic that was very evident when putting together a list of invitees for the meeting was that the ultrahigh vacuum/atomic resolution scanning probe community in the UK is rather skewed towards blokes of “a certain age” (and that most definitely includes me.) It’s been suggested — by Eugenie Hunsicker (Loughborough) — that one way to attempt to address this would be to consider a collaborative incubator project, a scheme funded by the University of Bath’s Inclusion Matters programme. (Nottingham is also an Inclusion Matters grant-holder.) That is definitely a funding strategy I, for one, would like to pursue, alongside other EPSRC networking opportunities.
My slides for the meeting are embedded below. I will add the PowerPoint/pdf files for the other presenters, if and when I get permission, at the foot of this post. (Giovanni Costantini (Warwick) has already given me permission so his slides are the first there.) The core motivation for the meeting was to bring as many probe microscopists as possible together — and perhaps choosing the last working day before the start of the new academic year for many was slightly ill-advised from that perspective — to discuss strategies for ensuring that we don’t spend a lot of time “reinventing the wheel” when it comes to developing machine learning protocols. And our main objective is therefore to put together a UK-wide network of groups working on the machine-learning-enabled probe microscopy theme.
Despite the prevailing ‘wisdom’ in some deluded corners, the UK of course can’t stand alone, isolated and insular, when it comes to scientific research (or anything else for that matter.) Science is inherently international in scope, and the vast majority of research in this fair and sceptred isle thrives on collaborative activities with our colleagues outside the UK. When it comes to machine learning in scanning probe microscopy and nanoscience, in particular, we need to pay especially close attention to the exceptionally exciting and pioneering work being done by a number of groups across the world.
Some of those key groups are listed in the PowerPoint slides embedded above, including Bob Wolkow’s research team at the University of Alberta. Bob and his colleagues are very much setting the bar for the rest of us — particularly those of us who work extensively with silicon surfaces — when it comes to embedding machine learning in not just atomic resolution imaging but single atom manipulation. As Bob describes in this engaging TEDx video (uploaded just a few days ago), the STM (“See, Touch, Move”) is becoming ever more capable; one key advance that Bob’s group — in particular, PhD student researcher Taleana Huff and her colleagues — has made is the ability to repair/edit single atomic defects[3] :
Watch the video to get an insight into just how far the UoA team have pushed forward the state of the art in what is effectively 3D printing with atoms. Bob suggests that the latest advances from his group are potentially as disruptive as the transistor was to the vacuum tube. I’d cautiously agree with that statement, although moving from a UHV low temperature environment to the “big, bad world” outside the vacuum chamber is always going to be fraught with difficulty. I am looking forward immensely to spending a couple of weeks at UoA next year to learn more about the techniques pioneered there, thanks to funding from both UoA and Nottingham’s International Collaboration Fund.
I’ll provide updates on how the machine learning SPM network is progressing in future blog posts. For now, here are the slides from Giovanni, as promised above, and from Oli Gordon (Nottingham). (Oli is also pictured in the image that kicks off this post.)
Update 07:47 26/09/2019 It was hugely remiss of me not to highlight a very important upcoming (Jan 2020) conference in Kanazawa — The First International Conference on Big Data and Machine Learning in Microscopy. My sincere apologies to the organisers — sorry, Adam et al. — for not including this in the original post.
Presentations
Molecular-Scale Surface Analytics — Giovanni Costantini (Warwick)
Scanning Probe Tip State Recognition in Real-Time with Neural Networks — Oliver Gordon (Nottingham)
Machine Learning and (SI)STM — Peter Wahl (St. Andrews)
[1] …and how do we know what’s the right type of atomic resolution? That’s very much a moot point. Sometimes it’s whether the microscopist sees the same type of image as other groups have published previously. This is a slightly worrying way to do science.
[2] Note that we may not always want the highest possible spatial resolution. Different tip structures can have different densities of states, for one thing, and this can affect their ability to extract or move atoms (or molecules).
[3] Not that we’re bitter or anything, but an alumnus of the group, Peter Sharp, tried for very many months, years ago as part of his PhD, to get enough reproducibility to routinely “heal” single atom — more precisely, single dangling bond — defects in the manner that Taleana and her colleagues in Wolkow’s team have achieved. While Pete could definitely observe dangling bonds disappearing during a scan (see below — captured from Pete’s PhD thesis via my phone), which we interpreted as a transfer of a H atom from the tip, we could never quite get the transfer to happen reliably via chemomechanical force alone when we targetted a single dangling bond.