Putting A Spin On Machine Learning, Atom by Atom

点击翻译按钮获取中文摘要

There is nothing quite like the magic of magnets. And yet even Richard Feynman, an incredibly gifted science communicator, struggled to explain just how magnetism works. (The video in question is easily found on YouTube. Feynman's slight tetchiness with the interviewer who raises the subject of magnetic forces is not entirely unrelated to the difficulty in explaining their fundamental origin at a level that a non-physicist -- or, indeed, a physicist -- can readily grasp.) Scientists are now at the point, however, where not only can we measure forces on an atom-by-atom basis, but we can harness and exploit those self-same forces to manipulate magnetism right down to the atomic level (and beyond). The instrument that allows this exquisite level of control of magnetic forces is the scanning probe microscope. A technique that will shortly reach its fortieth birthday, probe microscopy is conceptually rather straight-forward -- its experimental realisation rather less so. An exceptionally sharp tip, terminated in a single atom or molecule, is brought extremely close to a surface such that the tip-surface separation is of the order of the diameter of an atom or less. This atomically sharp probe can then be used in a number of modes to explore, interrogate, and modify the underlying sample surface on an atom-by-atom basis. Some of the most exciting and ground-breaking science ever carried out has involved the scanning probe microscope's unparalleled ability to not only image, but manipulate, matter at the single atom level. Probe microscopes are not just limited to the imaging and control of atoms; they can go much further. With an appropriately modified tip apex, even the quantum mechanical spin of electrons -- which, ultimately, is the source of magnetism -- is detectable either via the tiny electrical current that flows between the probe and the sample, or, incredibly, via measurement of the minuscule magnetic force between single atoms. Just a couple of months ago (in Oct. 2019), Chris Lutz' group at the IBM Almaden Research Centre reported that they have achieved, in collaboration with researchers in Korea and Oxford, the most precise and coherent control of the spin state of individual atoms ever attempted with SPM. (It's worth noting that IBM is the birthplace of both the scanning probe microscope itself, which was invented by Binnig, Rohrer and co-workers in the Ruschlikon, Zurich research labs, and of SPM-driven single atom manipulation, due to the inspiring efforts of Don Eigler and colleagues at IBM Almaden.) But the deep, dark secret of the probe microscopist is that a very large percentage of their time is spent coercing and cajoling the probe into providing atomic resolution. Yet even that's not enough -- when that resolution is achieved, the microscopist very often has to maintain the ability to image, move, and spectroscopically interrogate single atoms at the same time, while always being on the look-out for tip-derived artefacts. The component at the core of probe microscopy -- the probe itself -- therefore represents a major, and infuriating, bottleneck in the technique. This project integrates artificial intelligence, surface science, and nanoscience to take the pain out of probe microscopy. We will develop a machine learning framework that, in essence, "auto focuses" a probe microscope and then takes the SPM to the point where it can learn how to build magnetic nanostructures atom-by-atom and spin-by-spin. By itself. This AI-enabled probe microscope will then be used to carry out a programme of exceptionally challenging experiments whose common theme is the control of magnetism at the most fundamental levels: single domains, single molecules, single atoms, and single spins.