Underwater sensor networks send data with sound because seawater absorbs radio. Sound travels slowly, bounces off the surface and seabed, and fades unpredictably, so the protocols that decide how much error-correcting redundancy to attach to each packet matter a great deal. Almost all of these protocols are designed and scored in network simulators that model the ocean as smooth distance-dependent attenuation plus random noise. Physical-layer researchers have already shown that such idealized models misrepresent real channels. This proposal asks a narrow, testable question: when a transport protocol picks its error-correction strategy based on a simulator, does it pick the same strategy that would win on a channel actually recorded at sea? If not, how much delivery reliability and energy does that cost?
We answer by driving a fixed modem receiver and a family of coding strategies with field-recorded channel impulse responses, which are measured snapshots of how the water smeared and delayed a transmitted pulse. The same strategies are also run in the idealized simulator. We then compare which strategy each environment ranks best. The second contribution is a calibration test. We fit a small model of bursty, time-correlated packet loss to one recording site and check whether it restores correct decisions at a different site. This tests whether cheap calibration can stand in for sea trials, or whether it only overfits to one patch of water.
Core Claim
Transport-layer conclusions drawn under smooth attenuation and independent noise are not just optimistic in absolute numbers. They can be wrong in ranking. Real acoustic channels lose packets in correlated bursts, so packet-level erasure codes and hybrid schemes should beat bit-level codes in regimes where an independent-error simulator says the opposite. We further claim that what matters for repairing these decisions is the temporal correlation of losses, not their average rate. If so, a two-parameter-per-regime burst model fitted to field data recovers most of the decision quality, and it transfers across sites better than tuning a ray-tracing simulator does. The first claim is falsified if simulator and replay rankings agree. The second is falsified if a mean-loss-only calibration does as well as the burst model.
Grounding in Prior Work
The motivating case is ARRTP, from Internode Distance-Based Redundancy Reliable Transport in Underwater Sensor Networks (also /work/internode-distance-based-redundancy-reliable-transport; Internode Distance-Based Redundancy Reliable Transport in Underwater Sensor Networks). ARRTP uses a distance-indexed lookup table to switch between three options: bit-level BCH codes, packet-level Reed–Solomon erasure codes, and a hybrid of the two. It was evaluated only in an extended NS-2.31 simulator with Thorp absorption and ambient-noise terms. We reuse its strategy space and its table-based decision rule as the object under test. Unlike that work, we never trust the simulator's verdict without checking it against measured channels.
Two physical-layer papers motivate the doubt. Underwater Acoustic Communication Channel Modeling Using Reservoir Computing trained models on Bellhop ray-tracing output and found they predicted poorly on Lake Tahoe recordings without fine-tuning on measured data. Channel Prediction for Underwater Acoustic Communication: A Review and Performance Evaluation of Algorithms benchmarked predictors on recorded AUVFest07 and Watermark NCS1 impulse responses and found behavior that changed sharply between calm and rough seas. Both papers stop at the physical layer. Neither asks whether the channel mismatch they observed flips higher-layer decisions, and that question is ours.
Two other papers constrain the design. The concern raised by Traversing the Reality Gap via Simulator Tuning and related sim-to-real work is that tuning a simulator to one site can overfit it to that site. This is why our calibration is always scored across sites, never within one. Developing an Underwater Network of Ocean Observation Systems With Digital Twin Prototypes - A Field Report From the Baltic Sea reported that short 64-byte instant messages were more reliable than higher-rate burst mode in the field. That is anecdotal but direct evidence that real channels punish long contiguous transmissions in ways an independent-error model would not predict.
The closest three-direction work, System-Level Optimization of AUV Swarm Control and Perception, validates with digital twins and bench-top modem emulators, but its emulated channels are synthetic. Our approach differs because the channel itself comes from field recordings, and the evaluated object is the coding decision rather than a swarm controller.
Research Program
The replay harness. We fix one simple, reproducible single-carrier receiver: a decision-feedback equalizer with standard Doppler compensation. We pass packets through time sequences of measured impulse responses from AUVFest07 (calm and rough segments), Watermark NCS1, and the tank and Lake Tahoe recordings from the reservoir-computing study. The receiver is deliberately ordinary, because the goal is to measure the channel's effect on protocol decisions, not to win at the physical layer.
Emulating distance. The recordings are made at fixed ranges, so we emulate ARRTP's distance axis by shifting signal-to-noise ratio along the Thorp-plus-spreading curve. This preserves each recording's fading and multipath structure. It is a stated limitation: changing range in reality also changes the multipath geometry, which a fixed recording cannot reproduce.
Running the strategies. Each coding strategy runs over the resulting bit and packet error streams. These are BCH, Reed–Solomon at several redundancy levels, hybrid RS–BCH, and ARRTP's adaptive table. For each strategy we record message delivery probability and energy per delivered bit, using ARRTP's energy model.
The idealized twin. The same strategies run through the NS-2.31-style model at matched average signal-to-noise ratio.
Calibration. We fit a two-state burst-loss model, which alternates between a "good" and a "bad" channel state with fitted switching rates, to packet outcomes from one site. We then rebuild the decision table under that model and evaluate it on the other sites.
Decision-quality metrics. The primary quantity is decision regret: the energy and reliability lost by using the strategy a given model recommends instead of the one that is actually best on the replayed channel. We also report rank agreement between simulator and replay, measured by Kendall's tau over strategies at each signal-to-noise point.
Baselines
The naive reference point is the do-nothing evaluation: the idealized Thorp-plus-ambient-noise simulator used in the ARRTP paper, taken at face value. In that simulator ARRTP reports a transmission success probability of at least 99% while using 12% to 35% less energy than fixed Reed–Solomon schemes (ARRTP). We re-run this simulator ourselves with identical strategies and energy accounting, so the replay comparison differs in only one variable. A replay delivery rate well below 99%, or an energy advantage over fixed Reed–Solomon that shrinks or reverses, would be our first headline result.
The strongest existing method is ARRTP's distance-strategy table itself. It is the only adaptive redundancy scheme in the file with reported numbers, and it already beats fixed coding in its own evaluation. We run it unmodified on replayed channels alongside fixed BCH, fixed Reed–Solomon, and hybrid coding at matched signal-to-noise ratio and identical packet sizes. A calibrated table "beats" ARRTP if its average regret on held-out sites is lower and its delivery probability is at least as high, with confidence intervals from bootstrapping over channel segments.
The physics-based competitor is a Bellhop-tuned simulator: ray-tracing driven by each site's sound-speed profile where available. It receives the same fitting data as our burst model and is scored on the same held-out sites. No reported transport-level result was found on Lacuna for Bellhop-driven protocol evaluation. The reservoir-computing paper reports only that Bellhop-trained channel models transferred poorly to Lake Tahoe without fine-tuning.
The key ablation is a mean-loss-only calibration. It matches each site's average packet loss rate but keeps losses independent across packets. This isolates temporal correlation as the active ingredient. If it does as well as the burst model, our mechanism claim is wrong.
For the field datasets themselves, no reported numerical result was found on Lacuna for AUVFest07 or Watermark NCS1 in the summary of the channel-prediction review, so we establish those numbers fresh.
Why not just…?
Why not just run a sea trial? A sea trial gives real channels, but only at one site, one season, and one sea state, at high vessel cost. The Baltic field report shows how much can be learned from such a trial, yet it yields only qualitative protocol conclusions. Replay across five recorded environments gives cross-site evidence that one trial cannot. We still treat a short physical run as the final check, described below, rather than claiming replay replaces water.
Why not just tune Bellhop or another physics simulator to each site? This is the real threat. A well-tuned propagation model might capture the multipath that drives burst losses. The reservoir-computing results suggest raw Bellhop transfer is poor, but they concern channel prediction, not coding decisions. We therefore include the tuned simulator as a full baseline under identical fitting data. If it matches the burst model on held-out sites, we will report that simple physics tuning suffices.
Why not just add interleaving and keep the simulator? Interleaving turns bursts into scattered errors, which might make the independent-error model approximately correct. We include interleaved variants of every strategy in the replay. If interleaving collapses the simulator-versus-replay regret to near zero, the practical conclusion is "always interleave," and that is a useful negative result. The cost is latency over multi-second links, which we report alongside.
Decisive Checks
The first check is rank reversal. In at least one of the recorded regimes, most plausibly AUVFest07 rough sea, the simulator's chosen strategy must have significantly nonzero regret on replay. If rankings agree everywhere, the simulator is adequate for this decision and the project ends with that finding.
The second check is cross-site transfer. We calibrate on AUVFest07 and test on Watermark NCS1 and Lake Tahoe, then repeat with the sites rotated. The burst-calibrated table must beat both the unmodified ARRTP table and the mean-loss ablation on held-out regret.
The third check is a short physical run. A short tank or pool transmission with a commercial acoustic modem will compare measured delivery against the replay prediction for the two or three strategies that matter. The check passes if the replay prediction falls closer to the physical measurement than the idealized simulator's prediction does. This last step is the one place we test whether the replay harness itself has a reality gap.
References
- Internode Distance-Based Redundancy Reliable Transport in Underwater Sensor Networks
- Underwater Acoustic Communication Channel Modeling Using Reservoir Computing
- Channel Prediction for Underwater Acoustic Communication: A Review and Performance Evaluation of Algorithms
- Traversing the Reality Gap via Simulator Tuning
- Developing an Underwater Network of Ocean Observation Systems with Digital Twin Prototypes - A Field Report from the Baltic Sea
- System-Level Optimization of AUV Swarm Control and Perception: An Energy-Aware Federated Meta-Transfer Learning Framework with Digital Twin Validation