Skip to main content
Ocean Acoustics → Investigation 02

Which Input Drives the Regulatory Decision?

OAT screening says bottom properties dominate the shutdown zone. But that analysis holds source level and mitigation fixed. A full Sobol variance decomposition across 28,672 evaluations reveals: 56% of take estimate variance comes from interactions OAT cannot see.

28,672
Evaluations
6
Inputs Varied
56%
From Interactions
Three Sites

Same Model, Different Environments

We applied the same RAM PE model at three active or planned offshore wind sites. Each has different water depth, bottom composition, and distance from the NARW migration corridor. If the sensitivity ranking changes across sites, site-specific surveys are essential. If the ranking is stable, the same survey strategy works everywhere.

Site Location Water Depth Bottom Type Baseline Zone (160 dB)
Vineyard Wind (VW) S. Massachusetts 37 m Sand / glacial till 5.2 km
CVOW Virginia Beach 25 m Sand / clay 6.8 km
Revolution Wind (RevWind) Rhode Island 37 m Sand / gravel 4.9 km

The shallower site (CVOW, 25 m) produces the largest baseline shutdown zone. In shallow water, sound interacts more strongly with the bottom, and the lower absorption means energy propagates farther. This is consistent with Investigation 01’s finding that bottom interaction dominates.

Sound Speed Profiles — January vs July (WOA2023)

Sound speed from WOA2023 monthly T/S via Mackenzie (1981). January (winter, well-mixed) vs July (summer, stratified thermocline). All three sites show 40–50 m/s seasonal variation at the surface but converge below 30 m.

Sensitivity Ranking

The Tornado Chart

For each input, we varied it across its plausible range while holding all other inputs at their baseline values. The result is the spread in the 160 dB shutdown zone distance attributable to each input alone.

Shutdown Zone Sensitivity (160 dB threshold, all 3 sites combined)

Bottom attenuation dominates with 11 km of zone spread — more than all other inputs combined. The physical explanation is straightforward: in shallow water (<40 m), sound bounces off the bottom many times per kilometer. Small changes in how much energy each bounce absorbs compound rapidly with range.

Bottom type (sand vs. clay vs. gravel) contributes 4 km of spread, primarily through changes in the sound speed ratio at the water-sediment interface, which controls the critical angle for total internal reflection.

Seasonal sound speed profile contributes only 0.5 km. The summer thermocline creates a slight downward-refracting profile that increases bottom interaction, but the effect is small compared to the bottom properties themselves.

Bathymetry resolution (31 m CUDEM vs. 1 km interpolated) contributes only 0.2 km. The continental shelf is relatively flat in these lease areas; higher-resolution bathymetry doesn’t change the answer meaningfully.

One-at-a-time (OAT) sensitivity: each input varied across its plausible range while others held at baseline. OAT cannot capture interactions between inputs. Bottom attenuation and bottom type are physically correlated (both derive from sediment grain size via Hamilton’s regressions); their individual spreads cannot simply be summed. Combined across 3 sites (envelope of max minus min zone distance).

Global Sensitivity

What OAT Cannot See

OAT and Sobol point regulators at different surveys — and only one of them addresses what actually drives take.

OAT analysis identified bottom sediment properties as the dominant driver of shutdown zone size. Sobol analysis reveals that source level (S₁ = 0.74) and active mitigation (S₁ = 0.61) together explain most of the variance in marine mammal take — sediment properties explain roughly 5%.

These two analyses lead to opposite survey recommendations:

  • OAT-informed strategy: Commission a geotechnical survey to characterize bottom sediment. Addresses ~5% of take variance. Cost: $200K–$500K.
  • Sobol-informed strategy: Instrument the first few installation piles carefully to measure source levels and validate mitigation effectiveness. Addresses ~74% of take variance. Cost: comparable or lower.

OAT doesn't just give an incomplete picture — it points regulators at the wrong survey to commission.

OAT varies one input at a time, holding everything else fixed. That means it structurally cannot detect interactions — cases where the combination of two inputs matters more than either alone. We ran a Sobol variance decomposition (Saltelli sampling, 28,672 evaluations) across 6 inputs simultaneously to quantify both direct effects and interactions.

Input Range S1 (Direct) ST (Total) Interaction
Sobol Indices — Annual Take Estimate (6 inputs, 28,672 evaluations)

Source level (ST = 74%) and mitigation performance (ST = 61%) dominate the take estimate — not bottom properties (ST < 5%). But their first-order effects are only 23% and 15%. The gap is interaction: the S2 interaction between source level and mitigation alone accounts for ~25–30% of total variance. Physically, this makes sense: take scales with zone area (πr²), and the zone is the product of source-level range and mitigation reduction — a multiplicative interaction that OAT cannot detect.

The OAT tornado is not wrong — it’s incomplete. Within environmental inputs only, bottom properties do dominate. But the OAT held source level and mitigation at fixed assumed values (221.7 dB and 6 dB). When those assumptions are uncertain too, they dominate everything else. Sum(S1) = 0.44 means 56% of variance comes from interactions. The sensitivity ranking changes depending on what you hold fixed — and on the assumed prior ranges. If the pile design constrains source level to a narrow band (e.g., 218–224 dB for a known monopile), bottom properties would re-emerge as the leading driver. The Sobol analysis shows what matters when all inputs are uncertain; the OAT shows what matters once the engineering decisions are fixed.

Sobol analysis via SALib (Saltelli 2002). Uniform distributions on all 6 inputs. Nbase = 2,048, second-order indices computed. The model is a fast parametric surrogate: shutdown zone from Q2 RAM PE lookup (9 grain sizes, 4 freq, 4 azimuths) with analytical scaling for attenuation (zone × α−0.3) and source level (zone × 10ΔSL/17.5). Validated against 50 full RAM PE runs (400 propagation solves): R² = 0.925, median error 8.2%, positive bias +1.5 km (conservative — surrogate overpredicts zones). The surrogate degrades above SL > 225 dB where the analytical scaling overestimates range; the Sobol ranking is robust but exact index values carry ±10–15% uncertainty from surrogate error. Sobol indices are conditional on assumed prior ranges — narrowing the source level range (e.g., to a known pile design) would increase the relative importance of bottom properties.

Output Comparison

Different Decisions, Different Drivers

The Sobol ranking changes depending on which output matters. The shutdown zone (regulatory metric) is driven by source level (ST = 67%) and mitigation (ST = 46%), with moderate interactions. The masking zone (ecological metric) is almost entirely source level (ST = 97%) because bubble curtains are irrelevant at 50–80 km range. And NARW density only matters for the take estimate, not the zone distance.

Sobol Total-Order Indices (ST) — Three Output Metrics
Key Insight
The sensitivity ranking depends on the decision being supported. For regulatory compliance (take estimate), source level, mitigation, and density interact strongly. For ecological impact (masking), source level alone explains 94% of variance — no survey or mitigation strategy can reduce that uncertainty.

All three outputs computed from the same 28,672 surrogate evaluations (validated against 50 full RAM PE runs, R² = 0.925). ST confidence intervals < 0.07 for zone metrics; wider for annual take (up to 0.21) due to multiplicative scaling through area and density. Second-order indices confirm source level × mitigation interaction (S2 ≈ 0.15–0.31 for take and shutdown, ≈ 0 for masking zone; S2 has wider confidence bands than S1/ST at this sample size). Masking zone uses the same surrogate with range-scaled RAM PE propagation curve; uncertainty is higher at extreme ranges (>80 km).

Survey Priority

Where to Spend the Survey Budget

The sensitivity ranking translates directly into survey prioritization. If an input drives most of the uncertainty, the survey that resolves it has the highest value. If an input barely matters, the survey that measures it is wasted money.

Survey Resolves Est. Cost Zone Uncertainty Reduced Priority
Geotechnical (cores + CPT) Bottom attenuation + type ~$200K 15 km (bottom atten + type) HIGH
CTD / Glider Campaign Seasonal SSP ~$100K 0.5 km LOW
Multibeam Bathymetry Bathymetry resolution ~$50K 0.2 km LOW

Survey cost estimates are order-of-magnitude ranges based on published BOEM and IHA cost data for Atlantic OCS projects. Actual costs vary significantly with scope, vessel availability, and number of sites.

The Sobol analysis adds a new priority: characterize actual source level and mitigation performance through in-situ monitoring during initial piles. These two inputs explain 74% and 61% of take variance (total-order). A $200K geotech survey resolves both bottom parameters (ST < 5% each, partially overlapping) — important but far less than the 74% from source level. The most cost-effective “survey” may be careful measurement of the first few piles.

Depth Effect

Shallower Sites Show More Seasonal Sensitivity

While bottom properties dominate everywhere, the relative importance of seasonal sound speed varies with depth. At CVOW (25 m), the seasonal SSP spread is 3.2 km. At Vineyard Wind (37 m), it is only 0.5 km. The physics: shallower water has fewer propagating modes, and the thermocline occupies a larger fraction of the water column, so seasonal changes have a proportionally larger effect on the mode structure.

Site Depth Bottom Spread (km) Seasonal SSP Spread (km) SSP / Bottom Ratio
CVOW 25 m 14.2 3.2 23%
Vineyard Wind 37 m 10.8 0.5 5%
Revolution Wind 37 m 9.6 0.7 7%

This means CTD surveys may have value at shallower sites even though they are low priority at deeper ones. The survey recommendation is not universal — it depends on the site’s depth. But the geotech survey is high priority everywhere.

Finding
OAT screening correctly identifies bottom properties as the dominant environmental driver. But the full Sobol decomposition reveals that source level and mitigation — held fixed in OAT — explain 74% and 61% of take variance (total-order), with strong interactions (56% of variance). The optimal data-collection strategy is: (1) monitor actual source levels and mitigation performance during initial piles, (2) conduct geotechnical surveys, (3) skip CTD campaigns except at shallow sites.

OAT screening (3 sites, 4 environmental inputs) + Sobol global analysis (6 inputs, 28,672 evaluations, Saltelli sampling). 160 dB re 1 μPa SPL threshold. Ranges from literature and IHA applications.