Underwater navigation converts the time a ping takes to arrive into a distance, and that conversion depends on the speed of sound along the ray's path. Sound speed changes with depth, and it drifts over days as eddies, rotating blobs of warmer or colder water tens of kilometres wide, pass through. Today's high-precision tracking filters read their corrections from tables built once from a single stored sound speed profile, so the corrections go stale as the water changes. The filters also have no honest estimate of how wrong each range might be. We propose a small recurrent network that forecasts the evolving sound speed profile as a probability distribution. The recurrent network is an LSTM, a model that carries a memory state forward from one day to the next. The key design choice is how its forecasts are scored during training. The model is not graded only on sound speed at each depth. It is also graded on the distribution of acoustic travel times that its predicted profile implies for the ray geometries a navigation system actually uses.
The reason is that the downstream quantity, travel time, is roughly a weighted integral of slowness along a ray. Its uncertainty is therefore governed almost entirely by how errors at different depths co-vary. Per-depth losses ignore that co-variation. The sampling-based energy score, a common multivariate proper score, is known to be nearly blind to it (Scoring rule nets). A forecaster can look calibrated depth by depth and still badly misstate range uncertainty. Scoring directly in travel-time space targets exactly the correlations that matter. The resulting predicted variances can then be passed straight to a Kalman filter as per-ping measurement noise.
Core Claim
We claim three things. First, a recurrent sound speed forecaster trained with a proper scoring rule applied to travel-time projections of its predictive distribution will produce calibrated travel-time uncertainty. Second, it will do so where per-depth scoring, Gaussian likelihood training, and a linear Kalman filter on profile modes do not, and the gap should be largest on days when an eddy moves through. Third, passing these variances to an acoustic navigation filter reduces tail positioning error relative to the static-profile practice in A modified asynchronous sequential Kalman track fusion.
A scoring rule grades a predicted distribution against the outcome that actually occurred. It is called proper when the forecaster's best strategy is to report its true belief. The contribution here is not merely "an LSTM with a CRPS loss for sound speed". The contribution is choosing which quantities get scored, with that choice derived from the acoustics.
Mechanism
We represent each vertical profile by $K\approx 10$ empirical orthogonal function (EOF) coefficients. EOFs are the dominant depth-shapes of variation, the same vertical basis used in the RBF+EOF representation of Physics-Guided Reduced-Order Representation. A GRU or LSTM reads the recent history at a site: sparse past profile observations, standing in for occasional CTD or sound-velocity casts, together with daily surface temperature. It outputs a Gaussian mixture over the coefficients one to seven days ahead.
For a reference profile $c_0(z)$, we trace rays for a set of geometries $g$ (source depth, receiver depth, horizontal range) once, using Bellhop. A small profile perturbation $\delta c$ then changes travel time to first order as
$$t_g \approx t_g^{0} - \int_{\text{ray } g} \frac{\delta c(z)}{c_0(z)^2}\,ds \;=\; t_g^{0} + \mathbf{a}_g^{\top}\boldsymbol{\alpha},$$
where $\boldsymbol{\alpha}$ is the vector of EOF coefficients and $\mathbf{a}_g$ is a fixed sensitivity vector obtained by integrating each EOF along the reference ray. Because each travel time is a linear projection of $\boldsymbol{\alpha}$, the predicted mixture maps to a univariate Gaussian mixture for each $t_g$. The continuous ranked probability score (CRPS), which measures the squared gap between the predicted and observed cumulative distributions, has a closed form for Gaussian mixtures (Scoring rule nets). No sampling and no $O(D^3)$ covariance inversion is needed. The training loss is
$$\mathcal{L} = \sum_g w_g\,\mathrm{CRPS}\big(F_g,\ t_g^{\text{true}}\big) + \lambda \sum_{z}\mathrm{CRPS}\big(F_z,\ c^{\text{true}}(z)\big),$$
where $F_g$ is the predicted travel-time distribution for geometry $g$ and $w_g$ are geometry weights. $F_z$ is the predicted sound speed distribution at depth $z$, and $\lambda$ trades the acoustic term against a per-depth anchor. The weights $w_g$ can be tilted toward the geometries an operator cares about.
We are careful about propriety. The sum is proper, and it is strictly proper for each marginal it scores. It is not strictly proper for the full joint over depth unless the projection set is rich. We therefore also report a variant that adds random projections, and we treat any difference between the two variants as a finding.
The linearization is checked explicitly. On held-out days we recompute travel times with full Bellhop on the true profile and report the linearization error, alongside how often it exceeds the predicted spread.
Research Program
The primary data are the HYCOM South China Sea sound speed fields: 136 days, a $27\times27\times251$ grid at roughly 7.6 km horizontal spacing and 10 m in depth, with warm and cold eddies. These are the fields used by Striking The Right Balance and the RBF+EOF work. Each of the 729 water columns is one sequence.
We split the data in two ways. The temporal split holds out the final 30 days. The spatial split holds out a block of columns that an eddy crosses during the test period. We further tag test days as "eddy transit" or "quiet" according to the local sound speed anomaly, because the hypothesis predicts differences mainly on transit days.
The CMEMS reanalysis profiles and Shazikou casts used by Passive inversion of sound speed profile serve to fit the EOF basis on independent water, which tests whether the basis transfers. The measured sound velocity profiles and CTD casts from the 4500 m South China Sea AUV trial (Experimental Analysis of Deep-Sea AUV) are few in number. We use them only as an out-of-model sanity check on predicted spread, not for training.
For the downstream test, we simulate long-baseline navigation, meaning positioning from ranges to seabed transponders at known locations, using Bellhop on the true evolving HYCOM field. The filter receives either fixed measurement noise or the forecaster's per-ping travel-time variances.
As a secondary analysis in the spirit of NeuralHydrology, we ask whether any recurrent cell state correlates with local eddy anomaly. That study found individual LSTM cells tracking unobserved storage states. Such a cell would suggest the network is carrying physical memory rather than simply extrapolating.
Positioning Relative to Prior Work
The sound speed methods in the file motivate the gap. ME-SSPI inverts profiles fast but deterministically and statically (Passive inversion). The tensor network reconstructs 3D fields from sparse samples, but as point estimates on single snapshots (Striking The Right Balance). AS-KF-ESV corrects ray bending from static lookup tables whose memory grows from 1.4 MB to 68.7 MB at 5 m precision (A modified asynchronous sequential Kalman track fusion). Our sensitivity vectors are a compact alternative: one vector of length $K$ per geometry. The eddy studies in the Sea of Japan establish that perturbations reach about 2 m/s and refract sound substantially (Parameterization … Synoptic Eddy; Horizontal Refraction … Anticyclonic Eddy). That result is why a stored profile goes stale.
Three papers supply the scoring machinery. MVG-CRPS makes multivariate Gaussian CRPS tractable through whitening and shows that likelihood training inflates covariances under noise (MVG-CRPS). CCRPS gives the mixture closed forms and the negative result on energy-score correlation blindness (Scoring rule nets). ProbHardE2E trains sequence models with CRPS under hard physical constraints (End-to-End Probabilistic Framework). None of the three chooses the scored quantities from downstream physics, and that choice is what we add. The closest neighbour in navigation is the GNSS work that turns LSTM class probabilities into entropy-based measurement weights (Diffraction Identification). We replace that heuristic weighting with variances trained under a proper score in the units the filter consumes.
Baselines
The naive reference point is persistence. It holds the most recent observed profile fixed for the whole horizon, which is exactly what static-table filters assume, and its spread is the empirical spread of past forecast errors at that lead time. A climatological variant uses the training-period mean profile and variance. Any method that cannot beat persistence on travel-time CRPS at a three-day lead is not useful.
The strongest classical competitor is a linear-Gaussian Kalman filter on the same EOF coefficients, with autoregressive dynamics and process noise fitted by maximum likelihood on the training split. It is the natural choice of an oceanographer who already assimilates data. It produces a full covariance, and it maps into travel time just as easily through $\mathbf{a}_g$. It receives the same inputs, the same observation schedule, and the same basis, and its hyperparameters get the same search budget as the neural models. It is the baseline we most need to beat on eddy-transit days. It is also the one we most expect to tie on quiet days.
The neural competitors are the same recurrent network trained three other ways, with identical architecture, data, epochs, and tuning trials. The first uses a Gaussian likelihood on the coefficients. The second uses MVG-CRPS on the coefficients (MVG-CRPS). The third uses per-depth CRPS only, which is $\lambda$-only with $w_g=0$. The per-depth-only model is the ablation that isolates the key ingredient. If it matches the full loss on travel-time calibration, the acoustic projection is not doing the work. We also train a travel-time-only variant with $\lambda=0$ to see whether the anchor is needed.
For spatial reconstruction context, the tensor network reports an RMSE of 0.35 m/s at a 20% sampling ratio, against 0.52 m/s for Gaussian process regression, on the HYCOM South China Sea data (Striking The Right Balance). That figure is for snapshot reconstruction, not forecasting, so it bounds how accurate a good nowcast mean can be rather than serving as a direct target. We found no reported result on Lacuna for probabilistic temporal forecasting of sound speed profiles or travel times on HYCOM, so every forecasting number here will be new.
Beating the field means three things. The full loss must achieve lower mean travel-time CRPS than the Kalman-EOF baseline on eddy-transit test days, with a paired bootstrap interval excluding zero. Its 90% travel-time intervals must cover within a few points of 90%. In simulated navigation, its 95th-percentile position error must be smaller than both persistence and Kalman-EOF.
Why not just…?
Why not just run a Kalman filter on the EOF coefficients? This is the real threat. On slowly varying water, linear dynamics with Gaussian noise may be all that is needed. The case for a recurrent mixture model rests on eddy passages being abrupt, nonlinear, and producing skewed or bimodal short-term outcomes: will the eddy edge arrive tomorrow or not? The file notes that recurrent models with Gaussian heads miscalibrate under exactly such regime changes, but it offers no direct ocean evidence either way. The eddy-transit versus quiet split settles the question. If Kalman-EOF matches us on transit days as well, the conclusion is that the travel-time scoring idea should be applied to the Kalman filter's own tuning, by fitting its noise parameters with the travel-time CRPS, and the recurrent network is dropped. That variant is included so that this outcome remains a usable result.
Why not just train with Gaussian likelihood, or per-depth CRPS, and trust the covariance? Likelihood training is known to inflate covariances under noisy data. On electricity, the log score reached a CRPS of 0.0419 against 0.0249 for MVG-CRPS (MVG-CRPS). Per-depth scores say nothing about cross-depth correlation, and correlation is what sets travel-time variance. Both are run as baselines, and the per-depth-only model is the central ablation.
Why not just take an ensemble of deterministic forecasters, such as several tensor networks or MC-dropout nets, and use their spread? Ensemble spread reflects disagreement between models, not the intrinsic unpredictability of the water. Nothing in its training rewards calibrated travel-time spread. We include a five-member deterministic GRU ensemble at matched total compute. If its travel-time coverage is already near nominal, the case for training on a proper score weakens, and we will report that plainly.
Why not just recompute Bellhop on each forecast profile instead of linearizing? That is possible at evaluation time, and we do it to check the linearization. Inside training, however, it would require thousands of ray traces per batch and would not be differentiable. The sensitivity vectors make the travel-time score closed-form and cheap. The linearization-error report shows whether that shortcut is safe at the perturbation sizes seen in eddies, roughly 2 m/s.
Decisive Checks
Two outcomes decide the proposal. First, on eddy-transit days, the full loss must beat both Kalman-EOF and the per-depth-only ablation on travel-time CRPS and on coverage from the probability integral transform. We measure the latter by where each observed travel time falls in its predicted distribution, reporting coverage of central intervals. Second, feeding its per-ping variances into the navigation filter must reduce tail position error relative to fixed noise with the same mean forecast. That comparison separates the value of the uncertainty from the value of a better mean.
A clean negative is also informative. Suppose per-depth CRPS matches travel-time CRPS once $\lambda$ is tuned. Then the correlation-blindness argument does not bite in practice at this dimensionality ($K\approx10$), and the simpler loss should be preferred.
References
- Scoring Rule Nets: Beyond Mean Target Prediction in Multivariate Regression
- A modified asynchronous sequential Kalman track fusion based on sound propagation time
- Physics-Guided Reduced-Order Representation of Three-Dimensional Sound Speed Fields with Ocean Mesoscale Eddies
- Striking the right balance: Three-dimensional ocean sound speed field reconstruction using tensor neural networks
- Passive inversion of sound speed profile based on normal mode extraction of monochromatic signals
- Experimental Analysis of Deep-Sea AUV Based on Multi-Sensor Integrated Navigation and Positioning
- NeuralHydrology - Interpreting LSTMs in Hydrology
- The Parameterization of the Sound Speed Profile in the Sea of Japan and Its Perturbation Caused by a Synoptic Eddy
- On the Effect of Horizontal Refraction Caused by an Anticyclonic Eddy in the Case of Long-Range Sound Propagation in the Sea of Japan
- MVG-CRPS: A Robust Loss Function for Multivariate Probabilistic Forecasting
- End-to-End Probabilistic Framework for Learning with Hard Constraints
- Deep Learning-Based Diffraction Identification and Uncertainty-Aware Adaptive Weighting for GNSS Positioning in Occluded Environments