Frequently asked questions

The effect of anything you do in the ocean is a three-dimensional field that changes from hour to hour and extends well beyond the site. A sensor reports one value at one place and time. Reconstructing the field from a few points is underdetermined, and instrumenting tens of square kilometers for months costs more than the operation it is meant to check. A model that solves the physics reconstructs the field, and the sensors keep it validated.

A mechanisticmechanisticA model that solves physical laws from first principles, so it works in situations it was never trained on. model of your specific stretch of ocean, run continuously, informed by targeted in-situ and satellite observations, and reporting the quantities you need with an uncertainty range attached. It has three parts: a physics-resolving model, the instruments that validate it, and the reports that compare the ocean with your operation against the same ocean without it.

No. A statisticalstatisticalA model that learns patterns from historical data; fast where data is dense, unreliable outside it. or machine-learning model reproduces the examples it was trained on and is not reliable outside them. It also cannot produce a counterfactualcounterfactualThe same place and time simulated without the intervention. The difference between the run with it and the run without it is the attributable effect., because there are no data for the ocean without your intervention. atdepth solves the governing equations of motion and chemistry from first principles, so the model remains valid in conditions no one has measured, which is the normal situation for an ocean intervention. Machine learning is used for a specific task: calibrating uncertain parameters against observations.

Not by itself. An instrument reduces the uncertainty of a reported quantity only once a model uses its data: to set boundary conditionsboundary conditionsWhat you tell the model happens at the edges of the region you simulate. Poor ones corrupt the interior., to calibrate the parameters physics does not determine, and to test the prediction. The model in turn indicates where the next instrument should go. Modeling and monitoring are one system, and the twin is the two together.

The counterfactual is the same place and time simulated without your intervention. The model runs twice with identical forcing, once with the operation and once without, and the difference is the attributable effect, separated from tides, weather and natural variability, because the only thing that changed between the two runs is the operation. Every reported quantity, from tonnes of CO₂ to seabed area above a deposition threshold, is a difference of this kind.

Global products resolve the ocean in cells tens of kilometers across, and most regional models at about a kilometer. At that resolution a plume, a dosing field, a farm or an outfall does not exist in the model. atdepth uses those products for what they resolve well, the large-scale state and the boundary conditions, and nests a meter-scale model inside them for the site.

No. The discharge model is one output of a larger system. On the modeling side, the same run that resolves where the discharge goes also solves the biogeochemistry and returns the metrics the question needs: plume extent as a function of thresholds, fluxes across a boundary, and any other quantity a permit or a credit requires. On the monitoring side, atdepth builds and operates the Standard Monitoring Unit, a purpose-built station for coastal sites, and designs monitoring strategies and instrument systems for deep-ocean operations. Our monitoring solutions are informed by the models: we get the data we need and where we need it to inform and validate the digital twin. We also characterize your operations before field deployment through analytical, numerical and experimental studies to support design optimization and environmental permitting.

Two separate tests. Qualification checks the physics and numerics against cases with known answers. Validation compares the running twin with observations from buoys, ships and satellites at matched places and averaging windows, and reports the skillskillA quantified measure of how well a model matches reality: error, bias, and correlation.: error, bias and correlation. Validation continues for as long as the twin runs, and it covers the model output, not only the code.

The solver is open source under the MIT license and the physics is published. What is proprietary is the configuration, data ingestion and run infrastructure, which set cost and turnaround, not what the customer can see. Every study includes an implementation report that documents each equation, parameter and forcing dataset, and inputs and outputs are delivered in standard, self-describing formats, so a third party can reproduce the result.

Not on that evidence. A turbulent plume shifts within seconds, and no model reproduces one instant of it; neither does a second sensor a meter away. The comparison that constrains dilution and mixing is between the distribution of states the model produces over the near fieldnear fieldThe water within meters to a few hundred meters of a discharge, where the plume's own momentum and buoyancy drive the flow and vertical motion must be resolved explicitly. and the distribution the sensor samples over time. Point-by-point agreement on absolute pH, pCO₂ or sediment concentrations can be irrelevant to the reported quantity, and the report states which differences matter and which do not.

The regional model is validated where data exist, usually against moored current meters and the temperature and salinity profiles from the nearest surveys, and the same setup is then applied at the site. Far-fieldfar fieldEverything beyond the near field, where the ambient circulation carries and dilutes what was released and a regional model applies. transport is checked against semi-analytical envelopes, and the near field against dedicated measurements once the operation runs. The report states what the observations can and cannot support.

The error against the real ocean cannot be evaluated directly, so uncertainty is characterized by running an ensembleensembleMany simulations of the same case with the uncertain inputs varied across their plausible range. The spread of the results is the uncertainty.: many simulations over the plausible range of inputs and parameters, such as source rate, currents, settling velocity, diffusivity and stratification. The result is a spread, reported for example as the 10th, 50th and 90th percentiles, not a single number, and where a threshold is involved it is reported as a curve against the threshold, so the physics is separated from a limit the regulator may not have set yet. The ensemble also ranks which inputs matter, and that ranking determines where the next measurement goes.

It depends on the quantity. A mean can be estimated with far fewer members than the conservative percentile that carbon protocols require. atdepth sizes the ensemble to the statistic being reported and states the resulting standard error. Because the model runs on GPUs roughly a hundred times faster than legacy codes, ensembles of the required size are affordable.

In almost every study: the boundary conditions taken from the regional model, the source term (how much, where, how fast), and, for sediment, the particle properties. Numerical error is usually small compared with these. That is why the sensitivity study comes before the monitoring plan: the instruments go where the ranked uncertainties are.

Model results depend on the quality of the inputs. When customer-supplied records fail a mass-balance or calibration check, atdepth says so before using them and, where possible, builds a virtual counterfactual from what is known with confidence, such as the measured intervention chemistry and the feedstock stoichiometry. That construction conserves mass by design and can only understate the effect, so the reported figure stays defensible.

It does not predict the ocean perfectly; the ocean is chaotic, and the twin reports validated quantities with intervals, not certainties. It reports physical and chemical quantities: transport, concentration, deposition, carbonate chemistry, air–sea flux. Ecological significance, thresholds, cumulative-impact assessment and the permitting decision belong to the regulator and the operator.

What crosses the boundary of your operation: the discharge or dose rate and its chemistry or particle properties, positions and schedule, and any monitoring data you already hold. atdepth supplies the ocean data: bathymetrybathymetryThe shape of the seafloor; the depth map the ocean flows over., regional forcing, tides, rivers and weather come from public and satellite products. A short template lists the inputs, and provisional values can be used and marked as such until confirmed. Proprietary data are not required to start: the first pass runs on public data, and anything shared under restriction is flagged so it never reaches a published output.

A regional model of a new site is configured and validated in days rather than months, because it is built from physics and public data instead of years of site history. Monitoring units deploy in weeks; the lead time is set by hardware availability and shipping. The model can characterize a site, size a dose or design a monitoring layout before the first instrument is deployed, and most of a project can be set up before the customer confirms the final inputs.

Fewer than you expect, in places you might not have chosen. The model determines the layout: instruments go where the ranked uncertainties are and where the physics indicates the signal will pass, such as downstream and along slopes, plus a reference station outside the influence of the operation. Often one well-placed unit is the calibration reference for a large area.

A methodology document, an implementation report (setup, validation statistics, regional oceanography), periodic performance reports against the counterfactual with uncertainty ranges, and access to raw and processed monitoring data and key model outputs. Reports are written for the reader who has to act on them, whether that is an operations team, a credit buyer or a regulator.

No. An engagement delivers results, data and the documentation to verify and reproduce them; the modeling layer runs as a service. Completed simulations are static datasets with no retraining or upkeep, and outputs come in standard formats that your own team can use directly.

Operational. A near-field non-hydrostaticnon-hydrostaticSolves vertical motion explicitly, capturing fast processes like plumes and internal waves at fine scale. simulation nested in a regional far-field model is the baseline every engagement runs on, with the same software, validation and reporting infrastructure behind it. Development of new capabilitiesis funded separately through research programs and internal development funding; a customer engagement does not fund it. We are also happy to join collaborative research programs to develop and implement new capabilities.

We operate under two different models: monitoring-as-a-service and monitoring strategy design and integration. In monitoring-as-a-service, we own the hardware, deploy, recover, and maintain it. We charge a monthly fee for the service, and the data are delivered to the customer. The other option is that we design the monitoring strategy and, if needed, support the hardware integration work, but the hardware is purchased and owned by the customer. In most cases, data and model outputs are perpepetually licensed to the customer for their use in the context of the project and beyond. atdepth signs non-disclosure agreements, and customer names, sites and results do not appear in atdepth's material without written permission. Publicly funded research outputs are released openly under the funder's terms.

Engagements are scoped per project, so there is no established price list. The structure is the useful part: the model costs rises depending on the number of scenarios, the size of the ensemble, the complexity, and the required computational resources. Monitoring cost scales with instruments and number of units, and also varies with the deployment location. The model limits the monitoring to the minimum needed to resolve the question. In the case of MRV projects that lead to credit issuance, the cost is not tied in any way to the number of credits issued as a result nor a function of the credit price.

Non-hydrostaticnon-hydrostaticSolves vertical motion explicitly, capturing fast processes like plumes and internal waves at fine scale. near-field simulations at one to ten meters, regional hydrostatichydrostaticA simplification valid at large scales, where vertical pressure balances gravity. Fast, used for regional models. simulations at a hundred meters to a kilometer, nestednestingLinking a coarse model to a finer one, so each resolves the scales it should without solving everything at once. in one framework from about a meter to a thousand kilometers. Time steps adapt from seconds to hours. Physics and biogeochemistrybiogeochemistryThe chemistry and biology of the ocean: carbon, oxygen, nutrients, and the organisms that cycle them. run together on our own GPU servers or on professional cloud infrastructure.

The air–sea flux caused by the intervention is small compared with the natural variability driven by temperature, wind and biology, and most of the uptake completes over weeks, downstream, after the dosed water has left the site. A tonne of removal is a statement about a net flux integrated over a basin and a season. A sensor samples that field at one point; the model integrates it.

Because the removal is not realized until air–sea exchange occurs. Adding alkalinity or dissolved carbon to seawater changes the ocean's capacity to hold CO₂, but the atmospheric tonne appears only as additional air–sea exchange, reduced outgassing or increased uptake, wherever that water goes. Counting it requires following the water and the chemistry until equilibration.

Yes, for every marine pathway. Whether the process adds alkalinity, strips CO₂ from seawater, or grows and exports biomass, atmospheric removal happens only through air–sea exchange, and ocean storage is not the same as atmospheric removal. atdepth reports the two separately.

A single coefficient is valid only if the loss is the same everywhere and the system is near equilibrium, while the actual loss depends on the regional physics and chemistry: stratification, flushing, precipitation, and how long the water stays at the surface. Protocols that account for this treat a globally averaged factor as non-conservative. Resolving the ocean phase mechanistically replaces the factor with a site-specific, validated number and its uncertainty.

Yes, as a virtual experiment. The twin runs the planned intervention and its counterfactual under the site's real conditions, with the carbonate chemistry and air–sea flux resolved, and reports what a year of operation would remove and where. That is how a dose is sized, an outfall sited and a registry submission drafted before the first tonne is added.

The near-field model treats the intake as a virtual probe and the outfall as a source, resolves how much of the discharge returns to the intake under the site's currents and stratification, and reports realized removal against nominal capacity. Recirculation is a design variable, so the same run compares different intake and outfall placements to help optimize performance.

In the cases atdepth has modeled, regional uptake efficiency changes with the amount of alkalinity added are typically small. The practical limits is mineral precipitation near the outfall, set by the local saturation state, a near-field effect that coarse models do not resolve. The capacity–efficiency curve is set by discharge design and site, such as dilution, pipe geometry, exit velocity and proximity to a river mouth, so two plants of equal capacity can have different curves. That is a design question, and the twin quantifies it before construction. The presence of neighboring operations can also affect the efficiency, and the twin can quantify that effect.

Yes. Changes in total alkalinity, pH and aragonite and calcite saturation state are standard diagnostics of the coupled biogeochemistry, and acidification relief for fisheries or shellfish can be reported as a co-benefit, distinct from the creditable removal.

Export. A large fraction of a meadow's production leaves it as particles and dissolved carbon that cores and satellites do not measure, and export affects the accounting in both directions. So current stock-based protocols can under-credit or over-credit. atdepth resolves the fluxes, lateral export and air–sea exchange, so the service a meadow provides can be counted whether its biomass is changing or not.

A discharge plume at depth can span hundreds of cubic kilometers and persist for months. The most extensive campaigns have detected sediment several kilometers from the source, which is still only a few points at the edge of an evolving cloud. Any unsampled volume leaves the maximum extent unbounded. Monitoring constrains the inputs and tests the prediction; the validated model produces the regulated metrics.

Extent is never a single number; it depends on the threshold. atdepth reports seabed area above a deposition threshold, maximum reach and suspended volume above a concentration, each as a curve against the threshold and against the counterfactual seabed, so the statement can read "concentrations above X remain within Y of the operation". The colored region on a plot is where a chosen threshold is exceeded, not where the sediment stops.

Those tools are fast and simple, but they are based on integral plume theory with prescribed coefficients for buoyant jets in idealized geometry, treat sediment as a passive tracertracerA substance carried by the water (a dye, a chemical, dissolved carbon) whose transport the model follows. and the seabed as flat; in practice the coefficients are then tuned until compliance is reached, accumulating unquantified uncertainty. atdepth resolves the turbulence in three dimensions on the real bathymetry, coastline and ocean conditions, and passes the near-field result to a regional model with the physics intact. Resolving the physics changes the regulatory answer.

Before: near-seabed currents for weeks to months (five days is not enough), at least thirty continuous days where tides matter and a seasonal cycle where eddies matter, CTD profiles, and lab-analysis of sediment settling velocity distributions, which we can help with. During: turbidity sensors, current data, CTD profiles, settling velocity data. The model uses the moored velocities directly and indicates where the moorings should be.

The modeling engine runs ensemble simulations releasing thousands of Lagrangian particles to assess the connectivity of the operations site to the protected are of interest. The result is a probability distribution of the sediment reaching the area, and the report states the uncertainty of that probability. Avoid surprises and operational stoppages that cost time and money: the twin quantifies the risk before the operation starts.

Forecast oxygen, temperature, stratification and current velocityacross the whole farm extent and water depth days ahead; test a mitigation action, such as lowering pens or aerating, before committing to it; track the drift of sea lice and pathogens from neighboring sites; quantify the benthic waste footprint for compliance; and characterize a site before a lease. It runs in a cold fjord or warm openwater without requiring years of site history.

Yes, within the limits of the weather forecast that drives it: about ten days is the useful horizon for a coastal forecast, and the skill is reported alongside it. The twin also hindcasts past events at the scale you operate, which is how a mitigation plan gets tested before it is needed.

The tools in common use do not resolve the physics that sets a plume's fate and footprint, and project developers and regulatory agencies have told us they do not know how much of a study actually enters their decisions. atdepth delivers a satellite-driven regional model feeding a near-field turbulence-resolving model, run as an operational service, so siting alternatives are compared on resolved physics and every assumption is visible.

Any activity that disturbs the seabed, discharges to the water or changes the flow is the same physics: cable and foundation installation, dredging and disposal, scour, construction plumes, and turbine wakes acting on stratification. Regional models represent turbines through parameterizations that were never derived from near-field physics, which is why published wake studies disagree. atdepth runs the turbulence-resolving simulations those parameterizations should come from and returns the result in a form a regional modeler can use.

Anything the water carries can be followed through the resolved circulation: larvae with a settling schedule, a bloom with its growth and light limitation, a dissolved tracer. The limits are stated with the result. A velocity error of a few centimeters per second integrates to tens of kilometers over a month, so a connectivity fraction comes with one significant figure and an ensemble, not a false precision.

An MIT spinoff based in Boston that builds and operates Ocean Digital Twins for ocean interventions. Technology development is supported by the U.S. Department of Energy (ARPA-E), NASA and the Massachusetts Clean Energy Center. We are an ocean modeling and monitoring service company, and we work with project and infrastructure developers, operators, investors, government agencies, research institutions and non-profits.

On four continents, across ocean and river alkalinity enhancement, direct ocean capture, seagrass restoration and kelp MRV, aquaculture, offshore wind, sediment-plume assessment for deep-sea mining and dredging, and regional ocean intelligence and ocean modeling services. We also support design optimization, site selection, scalability assessments, operational decision making. Customers include top tier research institutions, government agencies,maritime operators, carbon-removal companies, infrastructure developers, and more. Names are shared with their permission, on request.

The physics runs on Oceananigans, a GPU-native, open-source ocean solver developed by the Climate Modeling Alliance and documented in dozens of peer-reviewed publications, coupled to an open biogeochemistry framework. atdepth currently hosts part of the continuous-integration testing for that code and builds its intervention physics, nesting and validation on top of it. Our team has contributed and continues to contribute to the solver's development, and the code is open source under the MIT license.

Yes. atdepth is paid to independently quantify what happens in the ocean, not to develop projects or sell credits. Reports state what the data support, including negative results.

Tell us the region, the operation and the question you need answered. We reply with what a twin can resolve for you, the inputs required from you, the outputs you will get, and how quickly.