Converting Between Light Polarizations: Speaker-Style Disentanglement for ARPES

Angle-resolved photoemission spectroscopy (ARPES) shines light on a crystal and records the energy and exit angle of the electrons it knocks out. From these measurements it builds a map of which electron energies are allowed at each momentum, which physicists call the band structure. The recorded image is not the band structure itself. It is the band structure multiplied by a "matrix element," a selection-rule factor that depends on the photon energy, the light polarization and the orbital character of each band. That factor can make a real band fade or vanish in one polarization and appear brightly in another. Voice conversion faces a problem with the same structure. A recording mixes what was said (content) with who said it (timbre), and the field has spent a decade learning to separate the two without letting timbre leak into the content representation.

We propose to treat each measurement condition (a polarization and photon energy pair) as a "speaker," and the intrinsic electronic structure as the "content." We would train a voice-conversion-style model to translate an ARPES measurement from one condition to another. ARPES offers something most voice conversion work lacks: parallel data. The same sample spot can be measured under several polarizations, which is equivalent to having the same sentence read by several speakers. The payoff we test is practical. A content representation that genuinely ignores the measurement condition should let a physicist recover bands that are suppressed in a single measurement. It should also make unsupervised phase mapping work in cases where current learned features fail.

Core Claim

Our central claim is that a condition-invariant content code improves band recovery from a single ARPES measurement. The code is learned from multi-polarization parallel ARPES data, with a structural bottleneck and a leakage penalty borrowed from voice conversion. Measured by recovered band positions and linewidths, a single measurement processed this way should approach what an experimentalist gets by measuring all conditions and combining them. A secondary claim is that a generative decoder is required to keep linewidths honest. We mean a rectified-flow decoder, which learns a straight-line transport from noise to data. A regression decoder trained with mean squared error would systematically broaden the peaks.

Both claims transfer specific lessons from voice conversion. The lessons are that content and style separate only under an enforced bottleneck, and that regression produces over-smoothed spectra. The transfer is not loose metaphor. Each lesson makes a falsifiable prediction in ARPES.

The closest prior work does none of this. A machine learning route between band mapping and band structure reconstructs band surfaces with a Markov random field, which is a probabilistic grid model with smoothness priors. It is seeded by density functional theory (DFT) calculations and requires the number of bands to be known in advance. It does not model the measurement condition at all; matrix-element suppression is simply something the DFT warm start must bridge. Transfer Learning Application of Self-supervised Learning in ARPES classifies micro-ARPES regions using ImageNet-pretrained features. It reaches 92.94% with labeled k-nearest-neighbors but only 50–60% with unsupervised clustering. We read that low unsupervised number as evidence that nuisance variation dominates the features, and that is the gap we target.

Research Program

The model follows the content/timbre/fusion design of ReFlow-VC. ReFlow-VC fuses a content stream, a pitch stream and a speaker embedding with cross-attention and gating, and decodes with a rectified flow. In our version, a content encoder reads an energy–momentum cut $I_c$ measured under condition $c$ and produces a code $z$. A condition encoder reads a short reference measurement under a target condition $c'$ and produces an embedding $s_{c'}$, playing the role of the speaker embedding. A rectified-flow decoder $v_\theta$ then generates the target-condition cut $I_{c'}$. We train it with the standard straight-path objective

$$\mathcal{L} = \mathbb{E}_{t, \epsilon}\,\big\| v_\theta(x_t, t \mid z, s_{c'}) - (I_{c'} - \epsilon) \big\|^2, \qquad x_t = t\,I_{c'} + (1-t)\,\epsilon,$$

where $\epsilon$ is Gaussian noise and $t \in [0,1]$ is the interpolation time.

Energy is fed to every module as an explicit coordinate referenced to the Fermi level (the boundary between occupied and empty states). This addresses the objection that image models wrongly treat a feature at the Fermi level the same as one 2 eV below it.

The content code is kept free of condition information in two ways. The first is a quantized bottleneck in the spirit of TVQVC, which forced VQ-VAE codes toward purely linguistic content. The second is an adversarial probe that tries to predict $c$ from $z$. Voice conversion found that bottleneck width trades content fidelity against leakage. The cross-lingual cloning benchmark in Toward Speaker-Preserving Arabic Speech shows a related trade-off between intelligibility and identity. SpeechSplit2.0 was designed to avoid hand-tuning bottleneck width. We sweep the width explicitly and report the trade-off curve rather than hiding it.

The intrinsic spectral function is read out as the decoder's output under a "neutral" condition embedding. In simulation the neutral condition is defined by setting the matrix element to a constant. On real data, a small calibration head maps $z$ to a neutral rendering, and that head is trained only on simulated cuts.

Training data come from two sources. The first is a simulator: tight-binding band models with self-energy broadening, a Fermi-Dirac cutoff, resolution broadening, Poisson counting noise, and dipole matrix elements computed for several polarizations and photon energies. The second is real multi-polarization ARPES data. Real data are the main logistical risk. The file documents the WSe₂ and Bi₂Te₂Se volumes of Xian et al. and the Fe₃Sn₂ micro-ARPES set of Ekahana et al., but it does not say whether multi-polarization parallel scans exist for them. We would therefore collect or obtain such scans from collaborating beamlines. If they cannot be obtained, we would restrict the real-data claims to the phase-mapping test, which needs no parallel data at inference.

Baselines

The naive reference is to take the raw single-condition cut and extract bands by peak-finding on its distribution curves. The distribution curves are intensity slices at fixed energy (momentum distribution curves) or at fixed momentum (energy distribution curves), fitted with Lorentzian line shapes. This is what most analyses do. Any suppressed band is simply missing from it.

The practical upper reference is the multi-condition sum. The same spot is measured under every available polarization, the intensities are added, and the same fitting is applied. This is what a careful experimentalist does with beamtime to spare. Our model sees only one condition at test time. We count it as beating the single-condition fit if it recovers suppressed bands. We count it as matching the sum if its band-position error falls within the noise of the multi-condition result.

The strongest existing method is the Markov random field of Xian et al., run with the authors' DFT warm start. It reports about 7 seconds per band on a GPU and 40–50 meV reconstruction error on synthetic benchmarks. We run it on identical simulated and real cuts with the same DFT input, and give it the true band count, which is an advantage our model does not receive. Beating it means lower band-position error than 40–50 meV-scale errors on matched synthetic data. More importantly, it means correct recovery in two regimes where DFT is deliberately wrong: rigidly shifted or mass-renormalized bands, and splittings absent from the DFT input. The second regime is modeled on altermagnetic splittings such as those in Large Band Splitting in g-Wave Altermagnet CrSb.

For phase mapping, the baseline is the DINO pipeline of Ekahana et al.. It reports 92.94% for k-nearest-neighbors with $k=20$, 80.36% for one-shot classification, and 50–60% for unsupervised k-means. We compare under the same splits.

Two ablations isolate the key ingredients. The first removes the bottleneck and adversary, giving an ordinary conditional autoencoder; this tests whether disentanglement matters. The second replaces the flow decoder with a mean-squared-error regression decoder, which tests the over-smoothing claim. All learned models get the same simulated corpus, the same GPU-hour budget, and the same number of hyperparameter trials.

Why not just…?

Why not just measure every polarization and sum them? When beamtime allows, this is a strong answer, and it is our upper reference rather than a straw man. It fails when time-resolved, spatially scanned or fragile samples permit only one condition per point, which is the regime of micro-ARPES. It also does nothing for features suppressed in all available conditions. If our single-condition model does not approach the summed result, the proposal has failed on its main claim, and the experiments will show this directly.

Why not just fit a physical forward model of matrix elements? For simple materials, a tight-binding model with computed dipole elements could explain the polarization dependence outright. This is a real threat on WSe₂-like systems. Its weakness is the same one Xian et al. report for DFT starts: correlated materials deviate from single-particle theory. Examples include the interlayer trion fluid in Electrically controlled interlayer trion fluid in electron-hole bilayers and the flat-band physics in Discovery of Van Hove Singularities. Our shifted-DFT and unseen-splitting tests are designed to separate the two approaches. If the forward-model fit wins even there, we would conclude that the learned style model adds nothing.

Why not just train a supervised denoising network on simulated data? A U-Net regressed onto clean simulated spectra is simpler. Voice conversion work, however, documents that mean-squared-error regression predicts the conditional mean and flattens spectral peaks: En-Najjary et al., Hanzlíček et al. and Mohammadi et al.. In ARPES that would inflate linewidths, which are exactly the quantity that encodes electron lifetimes. Our regression-decoder ablation tests this directly. If it shows no linewidth bias, we drop the flow decoder.

Proposed Experiments

The first decisive check is on simulated data, where ground truth is known. For bands whose matrix element is suppressed below the noise in the input condition, we measure band-position error in meV and recovery rate. We also measure linewidth bias, defined as the fitted width minus the true width. These are reported against all baselines across bottleneck widths. We also report probe accuracy for predicting the condition from $z$; chance-level accuracy is the target.

The second check is held-out conversion on real multi-polarization scans. We predict the unseen-condition cut from a single measured one and score its peak positions against the actual measurement. This is the ARPES analogue of the parallel-data evaluations of VCC 2020.

The third check uses Fe₃Sn₂ and asks whether k-means on our content codes rises clearly above the reported 50–60% unsupervised accuracy toward the 92.94% supervised figure. A clear rise would confirm that nuisance condition variation, not missing physics, caused the unsupervised failure.

The final check is the stress test against the Markov random field with wrong DFT input. Success requires that our model locate a splitting absent from the warm start, while the field model, given the wrong band count, does not.