The study, published in Frontiers in Marine Science, focuses on a specific and notoriously tricky region of the deep ocean known as the direct arrival zone, or DAZ. This is the area where a sound source’s energy reaches a receiver directly, without bouncing through the long refracted paths that dominate other parts of the deep ocean. When a source sits near the surface and a listening array hangs near the seafloor thousands of meters below, the received sound is shaped mainly by four paths: the direct wave, the wave reflected off the sea surface, the wave reflected off the seabed, and the wave that bounces off both. The coherent superposition of these arrivals creates alternating bands of high and low intensity in the vertical direction, a phenomenon the researchers call the vertical interference structure.

These interference stripes are not random. At very low frequencies, between 10 and 100 hertz, the spacing of the stripes is governed primarily by the time-delay difference between the direct wave and the bottom-reflected wave. Crucially, that spacing shifts noticeably when the target’s range changes but barely responds to changes in the target’s depth. At medium frequencies, between 1 and 10 kilohertz, the physics flips: the interference pattern, driven by the interplay of the direct and surface-reflected waves, is sensitive to both range and depth, with its period shortening as the source gets shallower. The research team, led by Junwei Zeng and Yixin Yang, realized that combining these two complementary sensitivities could yield a two-step localization scheme: use very low frequencies to nail down range, then use medium frequencies to resolve depth.

The central obstacle was computational. The traditional shortcut for predicting these patterns, the image source method, treats the ocean as if sound travels at a single constant speed, mirroring the source across the surface and the bottom to account for reflections. That approximation works tolerably close to the source, but in the far field of the direct arrival zone, defined in this study as ranges between 10 and 20 kilometers, the real ocean’s sound speed profile bends the rays so much that the image source method’s predictions fall apart. The team’s simulations showed correlation coefficients between predicted and true interference structures collapsing to near zero or even negative values at ranges beyond a few kilometers, meaning the predicted pattern bore essentially no resemblance to reality.

Full-fledged acoustic propagation models such as Bellhop or Kraken can handle the real sound speed profile, but at a punishing computational cost. In the team’s benchmarks, a ray-based approach that fed accurate arrival parameters into the interference equations took roughly six to ten minutes per scenario, far too slow for the rapid, repeated calculations that a matching-based localization method demands. The researchers needed something that combined the speed of the image source method with the physical fidelity of a full propagation model, and their answer was an elegant technique they call image-ray-family scanning, or IMRFS.

Instead of modeling the entire acoustic field, IMRFS traces families of sound rays numerically. Using Snell’s law and a set of ordinary differential equations solved with a standard numerical integrator, the method computes, for a thousand evenly sampled initial launch angles, exactly where each ray arrives, when it arrives, and how long its curved path through the ocean is. Reflections are handled with a clever mirroring trick: the source is reflected across the sea surface to account for surface bounces, and the receiving array is reflected across the seabed to account for bottom bounces, turning every multipath into an equivalent direct path through an extended, symmetric ocean. The result is a lookup table of arrival parameters that can be interpolated at any receiver depth, from which replica interference structures can be generated in seconds rather than minutes.

The performance gains were striking. In the far field of the direct arrival zone, the image-ray-family scanning approach produced interference patterns matching reference models with correlation coefficients as high as 0.87 at very low frequencies, while the old image source method’s coefficients plunged as low as negative 0.45. Computation times for the new method ranged from just five to nine seconds, compared with roughly 350 to 600 seconds for the accurate but slow ray-based alternative. That combination of speed and fidelity is what makes the subsequent matching procedure practical: the method can afford to generate thousands of candidate interference patterns and compare each one against what the hydrophone array actually measures.

The localization itself proceeds in two stages. First, the measured very-low-frequency interference structure from the target’s line spectrum is correlated against a bank of replica structures computed for scanning ranges from zero to 20 kilometers; the best match yields the estimated range. Because the very-low-frequency pattern is nearly blind to depth, this stage does not require any prior knowledge of how deep the target is, and simulations confirmed that even large errors in the assumed depth produced range errors of only about 0.2 kilometers. Second, with range fixed, the method computes medium-frequency replica patterns across scanning depths from zero to 200 meters and matches those against the measured medium-frequency structure to estimate depth. The two frequencies thus act as a coordinated pair, each covering the other’s weakness.

The robustness results are arguably as important as the headline accuracy numbers. The method tolerated deviations in assumed source depth, errors in the input range estimate, and variations in seafloor sound speed spanning 1500 to 2300 meters per second, with depth errors from bottom properties amounting to only about 3 meters. It performed well whenever the signal-to-noise ratio was 0 decibels or better, with range errors under 0.2 kilometers and depth errors around 4 meters, though performance degraded sharply in much noisier conditions. The team also mapped the method’s limits honestly: in certain destructive interference regions where acoustic energy cancels out, and at some specific source depths where unusual interference fringes appear, the technique struggles, and beyond about 19 kilometers the simple two-path physics begins to break down as additional reflected paths grow in strength.

For a field where localization errors of several kilometers and over a hundred meters in depth were common at long range, the simulated improvements, root-mean-square range errors dropping from 2.39 to 0.38 kilometers and depth errors from 57.4 to 2.5 meters, represent a genuine step change. The researchers emphasize that the method needs no precise knowledge of the seabed’s geoacoustic properties, no half-wavelength restrictions on the spacing of array elements, and no machine learning training data, making it attractive for real-world deployments on vertical arrays moored near the deep-ocean floor. If experimental trials at sea confirm what the simulations suggest, the invisible stripes of interference woven through the deep ocean could become one of the most reliable maps for finding things down there, whether the quarry is a whale, a research vessel’s pinger, or something nobody wants to admit is listening for.

Subject of Research: Underwater acoustic localization using vertical interference structures in the deep ocean

Article Title: A joint matching method for underwater target localization based on vertical interference structures at medium and very-low frequencies

Article References: Zeng, J., Yang, Y., Shen, H., Yu, M., Xu, S., & Ma, D. (2026). A joint matching method for underwater target localization based on vertical interference structures at medium and very-low frequencies. Frontiers in Marine Science, 13, Article 1947106. https://doi.org/10.3389/fmars.2026.1947106

Image Credits: AI Generated

DOI: 10.3389/fmars.2026.1947106

Keywords: underwater acoustics, deep ocean, target localization, interference structures, sound speed profile, ray theory, vertical array, range estimation, depth estimation, marine science, signal processing, Frontiers in Marine Science

The research team’s starting point is a candid critique of the two dominant approaches to range prediction. Purely physics-based models, built on equations of motion and battery electrochemistry, are interpretable and grounded in reality, but they struggle to capture the messy variability of real-world driving, from aggressive acceleration to stop-and-go traffic. Purely data-driven machine learning models, on the other hand, can fit historical driving data with impressive flexibility, yet they frequently produce unrealistic values, generalize poorly to conditions they have never seen, and offer almost no physical interpretation of their outputs. A neural network trained on summer highway driving, for instance, might confidently predict a range that is physically impossible in winter city traffic, and nothing in its architecture would flag the contradiction.

The authors’ solution is a Physics-Informed Neural Network, or PINN, an approach that embeds physical laws directly into the training process. Rather than allowing the network to learn freely from data alone, the framework adds constraints derived from energy conservation and the equations of motion that govern battery dynamics. During training, the model is penalized not only when its predictions deviate from observed data but also when its outputs violate physical consistency. The network must simultaneously predict the battery’s state of charge, the vehicle’s instantaneous power consumption, and the resulting driving range, and these outputs are constrained so that they remain mutually coherent. If the predicted power draw implies an energy drain that cannot be reconciled with the predicted change in state of charge, the training process pushes the model back toward physically plausible behavior.

This coupling of outputs is what distinguishes the framework from a conventional deep learning pipeline. In a standard network, state of charge, power and range would be treated as independent targets, and nothing would prevent the model from learning spurious correlations that happen to fit the training set. In the PINN formulation, the three quantities are bound together by the underlying energy balance of the vehicle. The battery’s stored energy, minus the integrated power consumed over the journey, must equal the remaining energy implied by the state of charge. By enforcing this relationship as part of the loss function, the researchers ensure that every prediction the model makes is one that the laws of thermodynamics would permit, dramatically reducing the risk of the unrealistic values that plague purely data-driven estimators.

To test the framework, the team turned to real-world data from the BMW i3, a compact electric vehicle whose driving ranges provided the empirical foundation for training and evaluation. The experiments were complemented by simulations conducted in SUMO, an open-source traffic simulation environment widely used in transportation research. This dual strategy allowed the researchers to assess the model both on genuine driving records and in controlled synthetic traffic scenarios, where factors such as route topology, congestion and driving style could be varied systematically. The combination of real telemetry and simulated traffic gave the evaluation a breadth that neither data source could provide alone, probing the model’s generalizability rather than merely its ability to memorize a single dataset.

The performance gains reported in the study are striking. The proposed PINN achieved a mean absolute error of just 3.39 percent in predicting the battery’s state of charge, and a mean absolute error of only 0.01 kilowatts in predicting power consumption, outperforming traditional neural network baselines on both metrics. In practical terms, a state of charge error of 3.39 percent means the driver’s estimate of remaining battery would typically be off by only a few percentage points, a margin small enough to make confident decisions about whether to continue driving or to stop and charge. The near-perfect power prediction accuracy is equally significant, because power consumption is the physical quantity from which range ultimately follows: knowing how much energy the vehicle draws per unit of distance is the essence of range estimation.

One of the most intriguing elements of the framework is its use of an attention mechanism, a technique borrowed from modern deep learning that allows a model to weigh the importance of different inputs dynamically rather than treating all information equally. In this application, the attention mechanism highlighted the decisive role of recent driving behavior in improving predictive performance. Intuitively, this makes sense: how a driver has behaved over the last few minutes is far more informative about immediate future energy consumption than a lifetime average of their habits. A driver who has just merged onto a motorway will burn energy at a very different rate than one crawling through urban congestion, and the attention mechanism lets the network focus on precisely the recent patterns that matter most for the next stretch of road.

The implications extend well beyond a single vehicle model. Range anxiety is consistently cited in surveys of consumer attitudes as a major deterrent to electric vehicle purchase, and the problem is partly psychological: drivers underestimate the reliability of their vehicles because the displayed range estimates themselves are unreliable. An estimator that is physically consistent, accurate to within a few percent, and sensitive to the driver’s actual recent behavior could transform the in-cab experience, replacing crude worst-case guesses with trustworthy, dynamically updated predictions. The authors argue that the fusion of domain expertise with deep learning makes their framework more robust, more physically meaningful, and more relevant to practical deployment than either paradigm alone.

The research also fits into a broader movement in machine learning known as scientific machine learning, in which physical constraints are used to regularize neural networks across domains from fluid dynamics to climate modeling. In transportation, the same philosophy is being applied to charging behavior prediction, smart charging optimization and energy consumption forecasting for fleets ranging from passenger cars to heavy-duty trucks. The present study’s demonstration that physics-informed constraints improve both accuracy and interpretability in range prediction adds a compelling data point to this trend, suggesting that the future of automotive AI lies not in bigger black boxes but in architectures that know the difference between what the data says and what physics allows.

Challenges remain before such systems reach production vehicles. The study relied on data from a single vehicle platform, and scaling the approach across different battery chemistries, vehicle classes and climates will require further validation. The authors note that the datasets generated and analyzed in the study are available from the corresponding author on reasonable request, which may facilitate replication and extension by other groups. Nevertheless, the core message of the research is clear and consequential: when neural networks are forced to obey the laws of physics, they stop hallucinating impossible ranges and start delivering estimates that drivers, and the engineers behind them, can rely on. For a technology whose mass adoption hinges on trust, that may prove to be the most important prediction of all.

Subject of Research: Physics-informed neural networks for electric vehicle driving range prediction

Article Title: Hybrid attention-enhanced physics-informed neural framework for accurate electric vehicle range prediction

Article References: Jeevakamal, K. R., Raj, R. D. A., Pallakonda, A., & Yanamala, R. M. R. (2026). Hybrid attention-enhanced physics-informed neural framework for accurate electric vehicle range prediction. Neural Computing and Applications, 38(17), Article 714. https://doi.org/10.1007/s00521-026-12400-9

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12400-9

Keywords: electric vehicles, physics-informed neural networks, range anxiety, state of charge, deep learning, attention mechanism, battery dynamics, energy consumption, BMW i3, SUMO simulation, range estimation, machine learning