MUTE-DSS: Digital-Twin Noise Mitigation System

• The paper introduces MUTE-DSS, a ROS2-centric digital twin system that integrates ship-noise modeling, 3D acoustic propagation, and a marine mammal distribution model to optimize voyage planning.
• It employs a two-stage optimization pipeline—using BIT* for routing and a genetic algorithm for speed profiling—that achieves significant noise reduction in sensitive marine environments.
• MUTE-DSS addresses underwater noise mitigation by combining digital twin technology with acoustically informed decision support to safeguard Southern resident killer whales.

Searching arXiv for the exact term and closely related papers to ground the article. MUTE-DSS, expanded as Mitigating Underwater Noise Transmission and Effects Decision Support System, is a digital-twin-based ship voyage planning system designed to minimize cumulative underwater radiated noise (URN) exposure to marine mammals, especially Southern resident killer whales (SRKWs), by jointly optimizing a ship’s route and speed profile under operational voyage constraints. It is described as a ROS2-centric framework that combines a ship-noise source model, 3D acoustic propagation, a data-informed SRKW distribution model, and a two-stage optimization pipeline for route and speed planning, with case studies in the corridor between the Strait of Georgia and the Strait of Juan de Fuca (Venkateshwaran et al., 3 Aug 2025).

1. Environmental problem, scope, and operating objective

MUTE-DSS is situated in the problem domain of ship underwater radiated noise in ecologically sensitive waters. The motivating context is the Salish Sea, where the Southern resident killer whale population is treated as especially vulnerable to shipping noise. The system is built around the operational observation that the effect of mitigation measures such as slowdowns and rerouting depends on vessel type, route, local bathymetry, propagation conditions, and where marine mammals are located, so mitigation cannot be evaluated adequately by a fixed rule alone (Venkateshwaran et al., 3 Aug 2025).

The paper frames the environmental concern in acoustic terms. It notes behavioral responses at received noise levels above 130 dB re 1 $\mu$Pa m (0.01–50 kHz) and Lombard-effect vocal compensation above 98 dB re 1 $\mu$Pa m (1–40 kHz). Because SRKWs depend on sound for communication and foraging, the system’s practical objective is to reduce modeled acoustic exposure while still respecting voyage requirements such as departure, destination, estimated time of arrival (ETA), and navigational constraints (Venkateshwaran et al., 3 Aug 2025).

The decision-support formulation is therefore not merely descriptive. MUTE-DSS is intended to recommend voyage plans that are acoustically informed. A plausible implication is that the framework treats underwater-noise mitigation as a constrained planning problem rather than as a single-parameter slowdown policy.

2. ROS2-centric digital-twin architecture and workflow

The system is presented as a digital twin (DT) in which a virtual vessel and marine environment are synchronized with live or simulated vessel and animal states. Architecturally, it is organized into a database unit, a modeling unit, an optimization unit, and a simulation and visualization unit. The paper also describes the workflow in three phases: case definition, model compilation, and dynamic runtime simulation and optimization (Venkateshwaran et al., 3 Aug 2025).

In the case definition phase, the operator specifies the departure and destination, ETA, permitted shipping lanes, exclusion zones, ship characteristics, environmental inputs, and the number of marine mammals to represent. Inputs include ship data such as engine specs, hull geometry, dimensions, AIS ship type, speed, and length; environmental data such as bathymetry, sound speed profile (SSP), and seabed properties; and marine-mammal data such as distribution/sighting information, dive-depth data, and movement assumptions (Venkateshwaran et al., 3 Aug 2025).

In the model compilation phase, the system prepares the computational components required for runtime use: a ship source spectrum model, a 3D acoustic propagation model, and a marine mammal distribution/trajectory model. The paper states that transmission loss is precomputed offline with Bellhop on HPC and later approximated for fast runtime access, whereas the ship source model is sufficiently light to be evaluated dynamically (Venkateshwaran et al., 3 Aug 2025).

In the dynamic runtime simulation and optimization phase, the simulation unit initializes a ship node, mammal nodes, a scheduler, and logging; the optimization unit computes a collision-free route and a speed profile; and the simulation and visualization subsystems update and render the resulting scenario. The paper describes this as a closed loop in which optimizer, ship state, mammal state, and acoustic evaluation continuously inform one another. This suggests a deployment trajectory toward onboard integration rather than a purely offline planning tool.

Phase	Main function	Principal outputs
Case definition	Specify voyage, ship, environment, mammals	Scenario inputs
Model compilation	Prepare source, propagation, and whale models	Runtime-ready models
Runtime simulation and optimization	Compute route, speed, noise footprint	DSS recommendations

3. Acoustic modeling and the marine-mammal representation

The acoustic core combines a near-field/source model and a far-field propagation model. The passive-sonar relation used in the system is

$\text{NL} = \text{NLS} - \text{TL},$

where $\text{NLS}$ is the near-field source level, $\text{TL}$ is transmission loss, and $\text{NL}$ is received noise level at the marine mammal. All quantities are expressed in dB ref $1~\mu$Pa (Venkateshwaran et al., 3 Aug 2025).

For source modeling, MUTE-DSS uses the JOMOPANS-ECHO (JE) model. The paper emphasizes that this model is based on 1,862 source level measurements collected by JASCO during the 2017 ECHO slowdown trial in Haro Strait and Strait of Georgia, and that it captures the speed dependence important for slowdown strategies. The source spectrum is evaluated at the center frequency of each one-third octave band from 12.5 Hz to $10^4$ Hz, with stated uncertainty of $\pm 6$ dB rms over 0.02–20 kHz (Venkateshwaran et al., 3 Aug 2025).

Propagation is modeled with 3D Bellhop ray tracing. The setup includes coherent transmission loss, a flat ocean surface, an acousto-elastic half-space bottom, 50 m spatial resolution, 360 azimuth angles, 360 elevation angles, a source depth fixed at 6 m, and receiver depths considered up to 100 m. Transmission loss was precomputed for 354 source locations along shipping lanes and over a 100 km radius around each source. To make runtime evaluation tractable, the Bellhop outputs were reduced to 10 principal components using PCA and then approximated by a Gaussian radial basis function (RBF) interpolant in a five-dimensional input space consisting of source latitude, source longitude, receiver latitude, receiver longitude, and receiver depth (Venkateshwaran et al., 3 Aug 2025).

The ecological component focuses exclusively on SRKWs. Initialization is data-informed through a kernel density estimator (KDE) fitted to sightings/encounters and maximum dive-depth data:

$$\mathrm{KDE}_h(\mathbf{x}) = \frac{1}{n h^d(n)} \sum_{i=1}^n \phi\left( \frac{\mathbf{x} - \mathbf{x}_i}{h(n)} \right), \quad \mathbf{x} \in \mathbb{R}^d.$$

After initialization, however, whale movement is simplified: each whale receives a random directional velocity vector with reasonable speed based on typical SRKW movement. The paper is explicit that more advanced stochastic trajectory forecasting models are discussed in the literature review but are not incorporated into MUTE-DSS (Venkateshwaran et al., 3 Aug 2025).

The cumulative exposure metric used in optimization is sound exposure level (SEL). For leg-wise received noise levels $\mu$0, the total SEL is written as

$\mu$1

with $\mu$2. This makes the optimization target voyage-long cumulative exposure rather than an instantaneous sound-pressure threshold (Venkateshwaran et al., 3 Aug 2025).

4. Two-stage optimization: route selection and adaptive speed profiling

The optimization pipeline is explicitly two-stage: first route planning with BIT*, then speed optimization with a genetic algorithm. The paper presents these stages as sequential rather than as a jointly converged bilevel or alternating optimization loop (Venkateshwaran et al., 3 Aug 2025).

For routing, the system uses Batch Informed Trees (BIT*) in the Open Motion Planning Library (OMPL) over an SE2StateSpace consisting of 2D position and yaw angle. Invalid states $\mu$3 include islands and invalid waters outside permitted lanes; valid states lie in $\mu$4. A route is a continuous path $\mu$5, and the route optimization problem is written as

$\mu$6

The route cost is approximated over waypoints, with local acoustic cost defined as

$\mu$7

Because minimizing negative transmission loss is equivalent to maximizing mean attenuation toward the whale nodes, the route planner seeks geometrically valid paths that exploit acoustic shielding by geography and bathymetry (Venkateshwaran et al., 3 Aug 2025).

Speed optimization is then performed along the selected route using a genetic algorithm implemented with Pymoo. The route is segmented into $\mu$8 sailing legs, with decision variables $\mu$9 for each leg. The total distance traveled is

$\text{NL} = \text{NLS} - \text{TL},$0

and the optimization problem is

$\text{NL} = \text{NLS} - \text{TL},$1

with $\text{NL} = \text{NLS} - \text{TL},$2 m. The reported GA settings include population size = 1000, adaptive crossover and mutation rates, and termination by objective-space tolerance or a predefined time window (Venkateshwaran et al., 3 Aug 2025).

A central design point is that the route stage optimizes where the ship should travel, while the speed stage optimizes how fast it should move along that path. This suggests a decomposition in which acoustically favorable geometry and acoustically favorable pacing are treated as distinct control variables.

5. Case studies and reported performance

The paper evaluates MUTE-DSS in voyages derived from AIS data for two ships departing from the Port of Vancouver: Star Kirkenes and Okiana. Two mammal scenarios are used. M1 is a simplified case with one stationary mammal node fixed at latitude 48.646343° N, longitude 123.313054° W, and depth 1 m. M2 is a more realistic case with five mammal nodes initialized using the KDE model and then moved dynamically with the simplified random-vector model (Venkateshwaran et al., 3 Aug 2025).

Case	AIS $\text{NL} = \text{NLS} - \text{TL},$3	Optimal $\text{NL} = \text{NLS} - \text{TL},$4	$\text{NL} = \text{NLS} - \text{TL},$5
C1M1	117.81	110.67	-7.14 dB
C1M2	118.55	113.65	-4.90 dB
C2M1	114.85	110.34	-4.51 dB
C2M2	116.40	115.48	-0.92 dB

The strongest reduction is reported for C1M1, where mean cumulative SEL falls from 117.81 dB to 110.67 dB, i.e. $\text{NL} = \text{NLS} - \text{TL},$6 dB. The paper interprets this as approximately an 80.68% reduction in linear exposure. It also reports a supplementary experiment in which the route is fixed to AIS and only speed is optimized; in that setting the reduction is 6.47 dB, compared with 7.14 dB for full optimization, implying that route modification contributed 0.67 dB in that scenario (Venkateshwaran et al., 3 Aug 2025).

In the more realistic dynamic setting C1M2, mean SEL is reduced by 4.90 dB, which the paper interprets as approximately a 67.6% reduction in linear exposure. The reductions are heterogeneous across individual whale nodes: one whale receives only 1.06 dB reduction, while another receives 8.68 dB reduction. This indicates that exposure reduction depends strongly on relative ship–whale geometry and on whether the route and speed profile can exploit local propagation shadowing (Venkateshwaran et al., 3 Aug 2025).

The most difficult case is C2M2, where mean SEL is reduced by only 0.92 dB. The paper reports that one whale, SRKW3, experienced 0.47 dB higher SEL than in the AIS voyage. The stated reason is that the tighter time and speed constraints left insufficient flexibility to maintain low speed in the critical Haro Strait region after earlier slowdowns had already been used to protect other whales (Venkateshwaran et al., 3 Aug 2025).

Across the reported examples, the qualitative behavior of the optimizer is consistent: it tends to make modest route changes, to apply selective slowdowns when whales are acoustically exposed, and to speed up in shadow zones created by islands or bathymetry. This suggests that the main value of the framework lies in geographically conditioned speed scheduling rather than in large-scale rerouting.

6. Limitations, deployment trajectory, and terminological scope

The paper gives a substantial list of limitations. Weather, waves, and currents are not modeled; uncertainty quantification is absent; fuel consumption, emissions, maritime fees, navigation risk, and comfort are not included as objectives; and vessel kinematics/dynamics and maneuvering constraints are not modeled. On the acoustic side, the ship is represented as a point source at 6 m depth, environmental parameters are fixed during each voyage, and the RBF surrogate can show artifacts in regions of steep transmission-loss variation. On the ecological side, whale motion after KDE initialization is simplified to random linear motion, and no explicit behavioral-response or threshold-risk model is added beyond exposure minimization (Venkateshwaran et al., 3 Aug 2025).

The deployment direction is nevertheless clear. The authors envision integration into an Integrated Bridge System (IBS), with onboard ship-state inputs, external passive acoustic monitoring, and a ROS2-based feedback loop between the physical vessel and the twin. This suggests that MUTE-DSS is intended as a voyage-planning and decision-support layer rather than as an autonomous navigation system (Venkateshwaran et al., 3 Aug 2025).

The term MUTE-DSS is specific to underwater-noise-aware ship voyage planning and should be distinguished from unrelated arXiv usages of MUTE or DSS. It is not the Mute deletion/suppression mode inside the Mirage and Mute Attack (M2A) framework for polyphonic sound event detection (Su et al., 2 Oct 2025); it is not MUTE = Message Unlearning for Targeted Efficiency in multi-agent reinforcement learning (Zuo et al., 3 Jul 2026); and it is not DSS = device-directed speech detection in multimodal voice-assistant systems such as M$\text{NL} = \text{NLS} - \text{TL},$7V (Wang et al., 2024). In the exact usage relevant here, MUTE-DSS denotes a digital-twin-based decision support system for reducing ship underwater radiated noise through acoustically informed route and speed optimization (Venkateshwaran et al., 3 Aug 2025).