Angle-resolved photoemission spectroscopy (ARPES) maps how electrons in a crystal are distributed in energy and momentum. It works by shining light on a sample and measuring the energy and exit angle of the electrons that fly out. Micro-focus ARPES narrows the light to a micron-sized spot, which makes the data four- or five-dimensional: two sample positions, angles, and energy. Raster-scanning that space costs days of synchrotron time. Autonomous micro-ARPES cuts the cost by letting a statistical model choose where to measure next. It fits a Gaussian process, a regression model that also reports its own uncertainty, and asks it which point would be most informative. Autonomous micro-focus angle-resolved photoemission spectroscopy showed that such a loop could reproduce a roughly 22,700-point raster map of Nb-doped Bi₂Se₃ with 300 to 600 measurements. That loop treats the sample stage in a simple way. It adds a penalty for long moves and otherwise assumes the sample sits exactly where the motor was told to go. Piezo stages in cold, ultra-high-vacuum cryostats do not behave that way. When a stage reverses direction it overshoots or undershoots (backlash), it keeps drifting after it stops (creep), and where it lands depends on the path it took (hysteresis). An active-learning loop that jumps around the sample reverses direction far more often than a raster scan does, so it is the regime where these errors should hurt most.
The core claim is that the autonomous loop should treat true sample position as an unknown that must be estimated from the motion history. The spectra themselves can serve as the position sensor. When the loop revisits a spot, the mismatch between the old and new spectra reveals how far the stage actually moved. A small history-dependent model trained on these self-calibrations can then correct commanded coordinates and report how uncertain its correction is. That uncertainty feeds into the Gaussian process and into the choice of the next move. Soft robotics supplies the method. There, recurrent networks whose internal state stands in for a material's deformation history beat static calibration by large margins. We carry that idea into a spectrometer, where it has not been applied.
Where this comes from
Two soft-sensing papers motivate the choice of a history-dependent model. Nonlinearity Compensation in A Multi-DoF Shoulder Sensing Exosuit For Real-Time Teleoperation used long short-term memory (LSTM) networks, recurrent networks that carry a running internal state. Its exosuit angle error was 2.51°/1.33° (azimuth/elevation), compared with 5.43°/3.66° for a feedforward network. A three-dimensional force estimation method for the cable-driven soft robot based on monocular images reported 0.84% mean relative force error, rising to 7.93% when the LSTM and motor-signal inputs were ablated. Both papers deal with viscoelastic polymers: rubbers whose chains take time to rearrange after they are deformed. Piezo stick-slip and backlash come from a different physical mechanism, so their numbers constrain our expectations rather than transfer to our setting. For that reason we test an LSTM against a classical hysteresis-operator model instead of assuming the LSTM wins.
KineSoft mapped sensor readings to shape using simulation plus alignment to real hardware. It supplies the sim-then-real training recipe we reuse. On the ARPES side, Transfer Learning Application of Self-supervised Learning in ARPES showed that image features from DINO, a model trained on natural photos without labels, can classify Fe₃Sn₂ micro-ARPES spectra into phases with 92.94% accuracy using nearest neighbours. We reuse those features as spectral fingerprints for matching revisited spots. That work treats each spectrum as a static image with a trusted position, which is the assumption we remove.
The hardware-side literature sets up our strongest competing argument. Ultra-sensitive, highly linear, and hysteresis-free strain sensors enabled by gradient stiffness sliding strategy removed hysteresis by redesigning the sensor itself. A Flexible and Robust Large Scale Capacitive Tactile System for Robots took the same approach by changing materials. Their view is that the fix belongs in the hardware. We take that seriously below.
Research Program
The first stage is measurement, because the premise could be wrong. On a micro-ARPES endstation we command the stage along the same kinds of back-and-forth paths the autonomous loop produces. Here "commanded" means the motor setpoint, while "actual" position is registered from a patterned calibration sample. Measuring both gives us the stage's error as a function of step size, direction reversals, and time since the last move. If that error stays well below the beam spot size at the temperatures used, the project stops here and reports a useful negative result.
In the second stage we build the self-calibrating position model. Its input is the command history. Its output is an estimate of the actual position together with an uncertainty. We train two versions for comparison. One is an LSTM in the style of ENNet. The other is a Prandtl–Ishlinskii hysteresis-operator model, a classical physics-based description of hysteresis. Both are fitted first on the calibration-sample data from stage one. They are then updated online from revisit residuals: the loop occasionally re-measures an earlier point, finds the position offset that best aligns DINO fingerprints and raw spectra between the two visits, and uses that offset as a training label.
In the third stage we plug the corrected positions into the Gaussian process from Ágústsson et al. Estimated position uncertainty enters the model as input noise. The loop then chooses its next point by weighing expected information against movement cost, as before, but now also against the growth in position uncertainty that a move would cause. Moves that reverse direction are penalised in proportion to that added uncertainty. We do not use a hand-set rule for reversals.
Baselines
The naive reference point is the published loop run unchanged: gpCAM with a mean-curvature feature and a distance-based movement cost, trusting commanded coordinates. The paper reports reaching a map faithful to the ~22,700-point raster with 300 to 600 points, a 40x to 75x reduction. It does not report map fidelity under measured stage error. We found no reported result on Lacuna for autonomous ARPES accuracy under stage hysteresis, so we must measure this baseline ourselves.
The strongest practical competitor is the same loop with the unidirectional-approach rule: every move finishes by approaching its target from the same side. This is the standard way experimentalists cancel backlash. The second competitor is a heteroscedastic Gaussian process, one whose noise level is allowed to vary across the map, so position error is absorbed as extra noise with no position model at all.
The ablations separate the key ingredient. One replaces the history-dependent model with a static per-direction offset. Another swaps the LSTM for the operator model. A third drops revisit-based online updating. An oracle run that uses calibration-sample positions, standing in for an ideal encoder, gives the upper bound.
All methods get the same sample, the same beamtime budget in wall-clock hours, including motion and settling, and the same Gaussian process hyperparameter search. Our method counts as winning if, within an equal time budget, it reduces the error in locating domain boundaries (in microns) and the map reconstruction error by a margin larger than the spread across repeated runs. It must win against both the unidirectional-approach baseline and the heteroscedastic baseline, not only against the naive loop.
Why not just…?
Why not just always approach each point from the same direction? For pure backlash this may work about as well as our model, and it is our most serious threat. It does nothing about creep that continues after the stage stops. It also costs an extra move whenever the next target lies on the "wrong" side, and the autonomous loop's jumpy paths make that frequent. The time-matched comparison settles the question. If unidirectional approach matches our method, we will report that the simple rule suffices.
Why not just install an interferometer or encoder on the stage? Where feasible, this is better, and our oracle run represents it. Retrofitting such sensors onto cryogenic UHV manipulators at existing beamlines is not always possible, and the hysteresis-free-by-design sensors from soft sensing have not been verified at cryogenic temperature and UHV. Our method needs no new hardware. Its value is measured by how much of the gap between the naive loop and the oracle it closes.
Why not just let the Gaussian process treat position error as noise? The heteroscedastic baseline tests exactly this. Noise that depends on the path taken is correlated with the scan history rather than random. We therefore expect a noise-only model to blur sharp boundaries instead of relocating them. The boundary-error metric will show whether that expectation holds.
Why not a static lookup table? The soft-sensing comparisons (5.43° vs 2.51°; 7.93% vs 0.84%) suggest that static models fail when the error depends on history. Piezo physics differs from polymer physics, so the static-offset ablation checks whether that lesson carries over.
Decisive Experiments
We first run the simulation study offline. We take ground-truth raster maps of Nb-doped Bi₂Se₃ from the Ágústsson et al. setting, obtained from the authors or re-acquired, and phase-labelled Fe₃Sn₂ maps. We corrupt their coordinates using stage error models fitted in stage one. We then replay every method against these corrupted maps, where the true answer is known.
The decisive check is then a live beamtime run on both materials that interleaves all methods under equal time budgets. We report boundary error, map reconstruction error, number of points, and total time. The DINO nearest-neighbour phase labels, at 92.94% accuracy on Fe₃Sn₂, set the classification reference against which we measure how much misregistration degrades the phase map.
The result that would falsify the idea is either of two outcomes: the stage error measured in stage one stays below the beam spot size, or unidirectional approach matches our method within run-to-run variation.
References
- Autonomous micro-focus angle-resolved photoemission spectroscopy
- Nonlinearity Compensation in a Multi-DoF Shoulder Sensing Exosuit for Real-Time Teleoperation
- A Three-Dimensional Force Estimation Method for the Cable-Driven Soft Robot Based on Monocular Images
- KineSoft: Learning Proprioceptive Manipulation Policies with Soft Robot Hands
- Transfer learning application of self-supervised learning in ARPES
- Ultra-sensitive, highly linear, and hysteresis-free strain sensors enabled by gradient stiffness sliding strategy
- A Flexible and Robust Large Scale Capacitive Tactile System for Robots