Heatwaves, cold snaps and pollution episodes often begin far upstream. A burst of tropical or continental heating launches a planetary-scale wave in the upper atmosphere (a Rossby wave), and that wave carries energy thousands of kilometers before it breaks and parks a high-pressure system over some unlucky region. Meteorologists trace this energy with wave activity flux, a vector field computed from wind and pressure anomalies that points the way wave energy travels. It marks sources where the field diverges and sinks where it converges. Today these flux maps are mostly read by eye, or summarized as scalar climate indices that throw away where the wave train sits and how it bends. We propose to treat each wave train as a biological specimen and measure its shape. Geometric morphometrics is the toolkit anatomists use to compare skulls or teeth after removing position, size and orientation. We would anchor its landmarks on physically defined points: the flux source, the flux sink, and points spaced along the flux path between them.
Our main claim can be tested and could fail. The downstream impact of a wave train depends on three things, and they can be separated: where the train sits (its longitude), how strong it is (its "size"), and its pure shape (curvature, tilt, and how far it spreads in latitude). Pure shape carries predictive information about downstream extremes that survives after conditioning on position, amplitude, and the standard climate indices. We test this with partial correlation, which measures the association between two variables after removing the linear effect of a chosen set of confounders. If shape adds nothing once position and amplitude are controlled for, the claim is refuted. That result would still be useful, because it would tell forecasters that a cheap Fourier phase and amplitude summary is enough.
Why this intersection, and what is actually new
The roles here are unequal. Wave activity flux supplies the object and the landmarks. Geometric morphometrics and functional alignment supply the machinery for putting different events into a shared coordinate system. Partial correlation is the test.
The physical motivation comes from three papers. Xue Han et al. show with ray tracing that small longitudinal shifts in tropical heating, which project onto nearly identical ENSO indices, send stationary Rossby waves along different paths. This is direct evidence that geometry matters and that scalar indices hide it. Evidence for preferred propagating terrestrial heatwave pathways due to Rossby wave activity builds a directed network of heatwave co-occurrence across Northern Hemisphere grid cells. It finds four preferred corridors (Asian, Western European, and two North American) where the horizontal Takaya-Nakamura flux is enhanced, and it reports that heatwave probability at downstream nodes within seven days is significantly higher when propagation is taken into account. That paper works on a fixed grid of nodes and never describes the wave train as a deformable object. Wenbin Kou et al. follow flux from North Atlantic blocking to its convergence zone over East Asia, again descriptively.
The method comes from the shared Procrustes core of the other two directions. Generalized Procrustes Analysis (GPA) removes translation, scale and rotation from landmark sets to leave "pure shape" (Sorting the flock). It uses sliding semi-landmarks along curves so that arbitrary point spacing does not inflate shape differences. Hyperalignment solves the same orthogonal Procrustes problem in high dimensions to match brains across people (A Model of Representational Spaces in Human Cortex). The main limitation this research file names for morphometrics is the homology bottleneck: nobody knows how to put corresponding landmarks on fluid patterns. Our answer is that the flux field supplies them. Its divergence maximum (the source) and convergence maximum (the sink) are physically homologous across events. The streamline connecting them is a natural curve for semi-landmarks. For a landmark-free variant we borrow fused unbalanced Gromov-Wasserstein alignment from Aligning individual brains with fused unbalanced Gromov Wasserstein. That method matches brains using both feature similarity and preserved internal distances, and its "unbalanced" relaxation tolerates regions that are missing or enlarged in one subject. The closest morphometric bridge, Automated morphological phenotyping using learned shape descriptors and functional maps, removes the need for manual landmarks, but it works only on static anatomical meshes and has no notion of direction.
One conceptual departure from standard morphometrics matters. In biology, position is a nuisance to throw away. In the atmosphere the background flow is not uniform in longitude: the jets and waveguides sit in particular places. Where a wave train sits therefore matters physically. We do not discard position and size. We keep them as measured covariates and use partial correlation to ask what each of position, size and shape explains that the others do not. This is the partial-correlation logic of Regional and seasonal partitioning of water and temperature controls on global land carbon uptake variability, which separated covarying water and temperature controls on carbon uptake. We apply it to the covarying geometric attributes of a single physical object.
The obvious nearby ideas are not what we propose, and each differs in a specific way. Causal discovery over climate fields, either with learned static spatial partitions (Learning a Spatial Partitioning and its Causal Relations from Temporal Data) or with Bayesian graphs over regional indices (Inferring the role of IPO phase dependencies), fixes the nodes in space. A meandering or refracting wave train therefore shows up there as edges switching on and off. Transformer forecasters (Deep Learning Meets Teleconnections, TelePiT, TeleViT) consume raw fields but never compute flux or expose an interpretable geometric variable. Our unit of analysis is the event-level wave train in shape space, which none of these use.
Research Program
Computing flux and extracting events. From ERA5 (1979 onward, regridded to 1.5°), we compute the horizontal Takaya-Nakamura flux at 300 hPa. It is built from 5-day-mean geopotential anomalies relative to a smoothed daily climatology, and that climatology also serves as the basic flow. We define a wave-train event as a connected region where the flux magnitude exceeds its 90th percentile and persists for at least four days. For each event we locate the source as the upstream maximum of flux divergence and the sink as the downstream maximum of convergence. We then integrate the flux streamline between them. We check event detection against the four heatwave corridors of Wang et al. and against the blocking cases of Kou et al.
Morphometric decomposition on the sphere. Each event contributes two fixed landmarks (source and sink) and roughly twenty semi-landmarks spaced by arc length along the streamline. The semi-landmarks slide along the curve's tangent during alignment. We work in three-dimensional Cartesian coordinates on the unit sphere. "Translation" is the rotation that brings the centroid of the configuration to a reference point, and we record it as a centroid longitude and latitude. "Size" is centroid size computed along the sphere's surface (geodesic), and we pair it with the mean flux magnitude to capture amplitude. "Orientation" is the residual rotation about the centroid axis, and we record it as a tilt angle. Because the variation between events may be large, we project the residual shape coordinates to a tangent space. We check the resulting distortion, one of the named limitations of geometric morphometrics, by comparing Procrustes distances with distances in the tangent space. Principal components of the tangent coordinates give interpretable shape modes such as arc curvature, meridional spread and zonal elongation.
Landmark-free variant. Some events have split or multiple sinks, where a two-landmark scheme would be arbitrary. For these we align each event's flux field to a set of template events using fused unbalanced Gromov-Wasserstein transport. The feature term compares the flux vector components, so direction enters the match. The geometry term preserves great-circle distances. The unbalanced penalty lets a lobe present in one event be absent in another. The transport plan induces correspondences, and we summarize the deformation relative to the template by the same position, size and shape split. We tune the entropic smoothing parameter on held-out years only, since the source paper reports alignment quality is sensitive to it.
Conditional attribution. For each downstream outcome $Y$, we estimate the unique association of the shape modes $S$ with $Y$ given a confounder set $Z$. $Z$ contains position, size, tilt, season, and the Niño 3.4, MJO and polar-vortex indices, all measured at or before event onset. The main statistic is the partial correlation $\rho_{YS\cdot Z}$. For binary outcomes we pair it with the likelihood-ratio gain from adding $S$ to a logistic model that already contains $Z$. We deliberately leave out of $Z$ anything measured after onset, such as downstream blocking indices. Those variables are plausible colliders or mediators, and conditioning on them would introduce Berkson's bias. Inference uses block bootstrap by year.
Baselines
The naive reference point is climatology plus persistence. This is the event-conditional base rate of each outcome for that season and corridor, combined with persistence of the current regime. For European regimes, Bommer et al. report that the NAO+ regime is a case where their models fail to beat persistence. No reported numerical skill for this baseline was found on Lacuna for the heatwave-corridor or regime tasks, so we recompute it on our splits.
The strongest statistical competitor is the index-plus-Fourier model. It is a logistic regression on the same climate indices plus the amplitude and phase of zonal wavenumbers 1 through 8 of 300 hPa meridional wind, averaged over the corridor's latitude band. Fourier phase already removes the translation that GPA removes, and Fourier amplitude already measures size. This baseline therefore attacks our claim directly. It gets the same event set, the same predictors measured at the same lead, and the same cross-validation folds. Beating it means a statistically significant increase in area under the precision-recall curve and in log-likelihood on held-out decades when shape modes are added.
The strongest learned competitor is the ViT-LSTM of Deep Learning Meets Teleconnections. It is a masked-autoencoder-pretrained vision transformer over full 2D fields feeding a recurrent network. It outperformed an index-based LSTM and outperformed ECMWF hindcasts beyond four weeks for Scandinavian Blocking and Atlantic Ridge. No reported numerical score for it was found on Lacuna. We retrain it on our splits using 300 hPa geopotential and flux fields as inputs, with a hyperparameter budget matched to our models. We do not expect shape features alone to beat it in raw skill. The fair test is whether our roughly ten interpretable shape modes, added to the index-plus-Fourier model, recover a stated large fraction of the ViT-LSTM's skill gain over indices. It is also whether they carry a nonzero partial association with outcomes after conditioning on the ViT-LSTM's own predicted probability. For context on pixel-level forecasting at these leads, ChaosBench reports a 44-day T850 RMSE of 3.56 K for ViT/ClimaX against 3.39 K for climatology. Direct field regression there is worse than doing nothing, which is why we predict event outcomes rather than fields.
Three ablations isolate the key ingredient. The first replaces flux-defined landmarks with extrema of geopotential anomalies, which tests whether homology anchored in flux matters. The second skips the Procrustes rotation and scaling and uses translated coordinates only, which tests whether separating size from shape matters. The third uses fixed-grid EOFs of the flux magnitude field in place of shape modes, which tests whether aligning events at all matters.
Why not just…?
Why not just use EOFs of geopotential or flux on a fixed grid? EOFs (the principal components of a field on a fixed grid) are standard and cheap. They are also a genuine threat whenever wave trains stay near fixed longitudes. The problem is that a wave train shifted in longitude splits its variance across several EOFs that differ in phase, so position and shape get mixed. The morphometrics literature documents the analogous failure: unregistered or linear measurements blur fine structure (Function-based Intersubject Alignment of Human Cortical Anatomy; A broad autism phenotype expressed in facial morphology). Neither paper, however, studies atmospheric fields. The EOF ablation settles the question directly. If fixed-grid EOFs match shape modes in held-out skill and in conditional association, alignment is unnecessary and we will say so.
Why not just summarize waves by Fourier phase and amplitude? This is the most serious alternative. Fourier summaries are global harmonics, whereas wave trains are localized packets with a source and a sink. Two packets with similar wavenumber spectra can arc along different paths, which is exactly the refraction effect Han et al. show for different ENSO flavors. Whether this matters for impacts is the empirical question, and the index-plus-Fourier baseline is built to answer it. The main hypothesis is the partial association of shape with outcomes given Fourier phase and amplitude. If that association is zero, the proposal's central claim fails.
Why not just feed raw fields to a transformer? Bommer et al. show that raw fields beat indices, so a transformer may well win on skill. What it cannot give is a statement like "curvature of the flux arc, not its position, raises Scandinavian heatwave risk". Its attention weights are unconditioned correlations between patches, which conflate direct pathways with covarying background modes, the same confounding that Ernst et al. exposed for summary-length effects in evaluation metrics. Testing the partial association of shape given the transformer's own output shows whether shape carries information the network misses. Recovering most of the transformer's gain with a few named variables would be a useful result in its own right.
Proposed Experiments
The first experiment uses the heatwave corridors. Using the Asian, Western European and two North American pathways of Wang et al., we predict heatwave onset at the pathway's downstream nodes within seven days of wave-train onset. Training uses 1979–2008 and testing uses 2009 onward, with metrics of area under the precision-recall curve, Brier score and log-likelihood. The decisive check is the sign, size and bootstrap interval of $\rho_{YS\cdot Z}$ when $Z$ includes Fourier phase and amplitude. It should be positive and stable across corridors. It should shrink toward zero under the geopotential-landmark ablation if flux-anchored homology is really doing the work.
The second experiment uses the European winter regimes. Scandinavian Blocking, Atlantic Ridge, NAO+ and NAO- serve as outcomes at leads of one to four weeks after the onset of upstream wave trains. NAO+ gets special attention because transformer models do not beat persistence there. A positive result on NAO+ from the shape modes would be the most informative single finding. The third experiment is a sanity check on physical consistency. Shape modes should relate systematically to the vertical wave coupling states described by Marlene Kretschmer et al., in which the stratosphere either absorbs or reflects upward wave flux, and to the jet position. If the leading shape modes correlate with nothing physical, they are probably artifacts of the event definition. We therefore repeat the whole pipeline under a flux-percentile threshold of 80 and 95, and under a four-day and a seven-day persistence window, and report how stable the modes are.
References
- Stability of ENSO teleconnections during the last millennium in CESM
- Evidence for preferred propagating terrestrial heatwave pathways due to Rossby wave activity
- Enhanced understanding of atmospheric blocking modulation on ozone dynamics within a high-resolution Earth system model
- Sorting the flock: Quantitative identification of sheep and goat from isolated third lower molars and mandibles through geometric morphometrics
- A Model of Representational Spaces in Human Cortex
- Aligning individual brains with fused unbalanced Gromov Wasserstein
- Automated morphological phenotyping using learned shape descriptors and functional maps: A novel approach to geometric morphometrics
- Regional and seasonal partitioning of water and temperature controls on global land carbon uptake variability
- Learning a Spatial Partitioning and its Causal Relations from Temporal Data
- Inferring the role of IPO phase dependencies and extratropical internal variability on the tropics
- Deep Learning Meets Teleconnections: Improving S2S Predictions for European Winter Weather
- Physics-Informed Teleconnection-Aware Transformer for Global Subseasonal-to-Seasonal Forecasting
- TeleViT: Teleconnection-driven Transformers Improve Subseasonal to Seasonal Wildfire Forecasting
- ChaosBench: A Multi-Channel, Physics-Based Benchmark for Subseasonal-to-Seasonal Climate Prediction
- Function-based Intersubject Alignment of Human Cortical Anatomy
- A broad autism phenotype expressed in facial morphology
- Re-Examining Summarization Evaluation across Multiple Quality Criteria
- The different stratospheric influence on cold-extremes in Eurasia and North America