Ice in the polar sheets is made of crystals that slowly rotate into a shared direction as the ice deforms. This alignment, called crystal orientation fabric, can make ice one to two orders of magnitude stiffer along the flow direction than randomly oriented ice. That is a larger effect than a 15–30°C temperature change (Gerber et al., 2023). The only non-destructive way to measure fabric continuously with depth is a ground-based radar that sends signals at several polarizations. It reads fabric from how the ice splits the signal into slow and fast directions (birefringence). The radar-derived profiles, however, come with no error bars. They systematically understate fabric strength compared with the few ice cores where crystals were measured directly (Ershadi et al., 2024). They also go blank in the bottom 10–30% of the ice column, where the signal degrades. Anyone who feeds these profiles into ice-flow models is using numbers of unknown reliability. There are only a handful of cores against which to check them.
We propose to treat each radar fabric profile as a smooth, non-linear curve of fabric versus depth, observed through a biased instrument whose bias has a physical cause we can model. The model pools all radar stations in one statistical fit and borrows strength across sites. It then calibrates its error bars using only the handful of co-located ice cores, with depth-blocked conformal prediction, a method that turns held-out errors into intervals with guaranteed coverage. The sharpest scientific bet is that much of the known radar underestimation comes from a specific assumption. Radar processing assumes the fabric's main axis points straight down. Flow modelling shows that axis can tilt by up to π/4 near ice divides, and ignoring the tilt underestimates horizontal eigenvalues by over 20% (Henry et al., 2025). If that bet holds, the bias can be predicted where no core exists, rather than only corrected where one does.
Core Claim
Radar-inverted fabric eigenvalues can be given depth-resolved, simultaneously valid uncertainty bands with honest coverage. These bands should extend into the radar-blind basal zone. Getting there requires three ingredients together. The first is a pooled non-linear depth-trajectory model on the constrained space of fabric eigenvalues, which are three non-negative numbers summing to one. The second is an observation model in which the radar sees only the horizontal projection of a possibly tilted fabric, plus a smooth residual bias. The third is a calibration step that uses the scarce cores as held-out blocks rather than as training labels. The claim is falsifiable in two ways. Adding the tilt term should shrink the unexplained radar–core gap at held-out sites. The calibrated bands should cover core measurements at close to the nominal rate while staying narrower than naive inflation.
Research Program
The latent quantity is the true eigenvalue triple $\lambda(z)$ at depth $z$ below a radar station $s$. We map this triple to unconstrained coordinates with a log-ratio transform, so the sum-to-one constraint holds automatically. We then model the transformed values as a population-level penalized spline plus a station-specific random smooth. This is the generalized additive mixed model structure used for irregularly sampled longitudinal data (Chen et al., 2024), with depth playing the role of time. Near the bed we impose a shape constraint. The degree of orientation must not decrease along the flow path, so under steady flow it must not decrease with depth. This condition comes from the stability constraint of Henry et al., 2025. We reuse it as prior knowledge, which governs how the curve is extrapolated into the unobserved basal zone. Boundary tails are restricted as in restricted cubic splines (Schilling et al., 2023) to avoid wild oscillation where data run out.
The observation model is where the physics enters. A nadir radar sees the fabric tensor projected onto the horizontal plane. If the dominant axis is tilted by angle $\theta(z)$ from vertical, the horizontal eigenvalue difference the radar recovers is a known geometric function of the true tensor and the tilt. We write
$$\Delta\lambda^{\text{radar}}_{H,s}(z) = P\big(\lambda_s(z), \theta_s(z)\big) + b(z) + \varepsilon_s(z).$$
Here $P$ is the horizontal projection. The term $b(z)$ is a smooth residual bias shared across sites, capturing whatever the tilt does not explain, such as inversion regularization. The term $\varepsilon_s(z)$ is spatially correlated noise.
The tilt $\theta_s(z)$ is latent. Its prior comes from diagnostic flowline fabric modelling of the same site, of the kind Henry et al., 2025 ran for Derwael Ice Rise. We will run the analogous Specfab-type calculation for the Hammarryggen, Dome C and EDML geometries. The magnitude of the noise is not guessed. At Hammarryggen, the 15 stations were measured in 2019 and again in 2020 (Ershadi et al., 2024). Differences between those repeat measurements act like disagreement between two annotators. This is the same trick Shi et al., 2024 use to supervise noise that comes from the measurement itself, here transplanted from medical image graders to repeat radar surveys.
We fit the model with Bayesian inference, which yields posterior bands. We do not trust those bands on their own. The final step is split conformal calibration. Each core site is held out as one contiguous depth block. The score is the maximum standardized deviation between the posterior and the core over the whole profile. Taking the maximum makes the resulting bands simultaneous over depth, not pointwise.
This replaces the wild bootstrap that Kong et al., 2024 use for simultaneous bands along 1D brain tracts. Their bootstrap is costly and assumes scalar, unconstrained responses. We borrow their idea of simultaneous bands along a continuous 1D domain but apply it to constrained tensor eigenvalues with a physical observation model. The honest limitation is that four core sites give very coarse conformal quantiles. So we additionally calibrate on synthetic sites, described in the experiments, and report both.
This program sits between existing work in a specific way. Ershadi et al., 2022 and Ershadi et al., 2024 supply the radar inversion. It is deterministic constrained least squares with a vertical-axis assumption and no error bars. We keep their inversion outputs as data and model what they get wrong. Matuk et al., 2020 give Bayesian uncertainty for orientation distributions, but for static metallurgical samples with no depth trajectory and no instrument bias. We share their goal of posterior uncertainty for fabric and add the continuous profile and the calibration against independent truth. Drews et al., 2015 model fabric transitions with depth in full-Stokes simulations but give no uncertainty. Their modelled girdle-to-single-maximum transition is exactly the kind of sharp inflection our spline must capture without smoothing it away.
Baselines
The naive reference point is the raw deterministic radar inversion of Ershadi et al., 2022, which has no correction and no uncertainty. We will attach the only error bar a practitioner could cheaply produce: the scatter of the Hammarryggen repeat surveys, applied uniformly with depth. Beating it means lower mean absolute error against held-out core eigenvalues and coverage closer to nominal. We found no reported result on Lacuna for a numerical radar-versus-core eigenvalue error for this inversion. The file only states that the underestimation is systematic. Our first deliverable is therefore to compute that number on identical depth grids.
The strongest competitor that is simple is a global linear recalibration. We fit radar equal to a times core plus b, using the core sites. We then apply it everywhere and inflate intervals by the residual spread. It receives the same leave-one-site-out protocol, the same data and the same conformal step. The comparison is fair because its only disadvantage is the absence of depth structure and tilt. Our model beats it only if held-out-site error drops. The gain should be concentrated at the flank and triple-junction stations, where tilt is predicted to be largest, and it should come with at least equal coverage.
The strongest existing uncertainty method is a pointwise Bayesian posterior from the pooled spline model without conformal calibration. This is the analogue of relying on model-based uncertainty. Perez et al., 2025 report that such model-based uncertainty (MC dropout, deep ensembles) collapses to 30–40% coverage against a 90–95% target under distribution shift. Meanwhile, their physics-residual conformal method holds 90–95%. We run the uncalibrated posterior with identical priors and sampling budget. Beating it means restoring nominal coverage at held-out sites without more than doubling band width.
Two ablations isolate the key ingredients. The first sets the tilt to zero everywhere, recovering the standard vertical assumption; this tests whether tilt explains the bias. The second removes the monotone orientation constraint below the radar cutoff; this tests whether the physical prior is what makes basal extrapolation honest. A third ablation fits each station independently rather than pooling.
Why not just…?
Why not just rescale the radar with a constant factor fitted at the cores? This is a real threat. If the underestimation reported by Ershadi et al., 2024 is a constant multiplicative shrinkage, the tilt term adds nothing. The linear-recalibration baseline above is designed to settle this. The physical reason to expect otherwise is that tilt varies strongly across a divide (Henry et al., 2025), so the bias should be larger on flanks than at summits. The Dome C transect, running from summit to flank, lets us test that pattern directly. If constant scaling wins at held-out sites, we will report that the bias is instrumental rather than geometric. We would then drop the tilt term and keep only the calibration and basal-extrapolation parts of the proposal.
Why not just run a full Bayesian inversion of the raw radar signal? A sampling-based inversion would give posteriors on eigenvalues at each depth, much as Matuk et al., 2020 do for orientation distributions. But it would inherit the same vertical-axis forward model. Its uncertainty would therefore be confidently centred on a biased answer, and nothing in it touches the core data. It also says nothing about the blank basal zone. Our uncalibrated-posterior baseline is a close proxy for this approach. The coverage comparison at held-out cores is what shows whether it falls short.
Why not just put physics-residual conformal bounds on a fabric evolution model? Perez et al., 2025 calibrate uncertainty without labels by measuring how badly predictions violate the governing equations. Those bounds live in the space of equation violation, not in eigenvalue units, and translating back is an unsolved inversion. More importantly, the radar–core gap is an observation error. A fabric model can satisfy its own equations perfectly and still disagree with the radar. We use their idea only as inspiration for making calibration cheap. The coverage target itself must come from cores.
Why not just use deep ensembles on the radar data? There are a few dozen radar profiles and four cores. Alonso et al., 2025 show that networks become confidently wrong under physical measurement corruption. Perez et al., 2025 show that ensemble coverage collapses under shift. With this little data, a flexible network has nothing to learn the bias from.
Proposed Experiments
The synthetic study comes first because it has complete ground truth. We take the 3D diagnostic fabric fields of Derwael Ice Rise from the Henry et al., 2025 modelling setup, which include tilts up to π/4. We pass them through the analytical radar reflection model of Ershadi et al., 2022, adding noise at the level measured from the Hammarryggen repeat surveys. We truncate the bottom 30% as observed at Hammarryggen. We then designate four random stations as "cores."
Three measurements on this synthetic data are decisive. The first is coverage of the simultaneous bands at non-core stations at 90% nominal, reported separately above and below the radar cutoff. The second is band width. The third is how much of the known over-20% horizontal-eigenvalue underestimation the tilt term recovers.
We also stress the model by giving it a deliberately wrong tilt prior, from a strain-rate-only rather than stress-weighted fabric model. This tests whether conformal calibration rescues coverage when the physics is misspecified.
The real-data study is leave-one-site-out across four sites. The Dome C transect has 20 stations across 36 km, and both Dome C and EDML have cores. Hammarryggen has 15 stations and a 263 m core (Ershadi et al., 2024). The NEGIS/EGRIP site has 6 radar locations and a deep core (Gerber et al., 2023). Its fast-flowing ice stream is a genuine distribution shift from the domes.
Metrics are mean absolute eigenvalue error against core thin sections, empirical coverage and width of the simultaneous bands, and a concordance check. The concordance check asks whether predicted uncertainty ranks the actual radar–core discrepancy across depth. It adapts the ranking logic of the label-free UQ scores of Pignet et al., 2025. We apply it in depth blocks because their i.i.d. assumption fails for autocorrelated profiles.
Two sanity anchors must be reproduced before anything else counts. The EDML flank value of Δλ = 0.265 and the Dome C summit value of 0.037 (Ershadi et al., 2022) must lie within our bands.
The proposal succeeds only under specific conditions. On synthetic data, the tilt term must recover a substantial fraction of the geometric bias, and conformal bands must hold nominal coverage where the uncalibrated posterior does not. On real data, the full model must beat global linear recalibration on held-out flank and ice-stream sites, not just on summits. With only four real sites, any real-data coverage claim is weak evidence. We will say so, and we will treat the synthetic results as the primary test of the calibration machinery.
References
- Crystal orientation fabric anisotropy causes directional hardening of the Northeast Greenland Ice Stream
- Investigating the dynamic history of a promontory ice rise using radar data
- Modelling the three-dimensional, diagnostic fabric anisotropy field of an ice rise
- Supervised Uncertainty Quantification for Segmentation with Multiple Annotations
- A functional nonlinear mixed effects modeling framework for longitudinal functional responses
- Polarimetric radar reveals the spatial distribution of ice fabric at domes and divides in East Antarctica
- Bayesian inference for polycrystalline materials
- Characterizing the glaciological conditions at Halvfarryggen ice dome, Dronning Maud Land, Antarctica
- Calibrated Physics-Informed Uncertainty Quantification
- ModelNet40-E: An Uncertainty-Aware Benchmark for Point Cloud Classification
- Legitimate ground-truth-free metrics for deep uncertainty classification scoring