AquAdapt AI · Cité Universitaire, Paris 2026
78 homes lost to coastal retreat. This was predictable.
The Scale
By 2050, nearly 1 billion people will live in low-elevation coastal zones exposed to rising seas.
IPCC SROCC, 2019
Nearly one in eight
people on Earth.
Coastal erosion is not a future risk. It is reshaping inhabited coastlines now — at a rate that outpaces the tools used to govern them.
What They See Today
Local governments, real estate developers, infrastructure planners get a spreadsheet. A rough estimate. Nothing parcel-level.
Slow
Built for specialists. Months of processing before outputs reach decision-makers.
Expensive
Requires expert teams to interpret. Inaccessible to most municipalities.
Static
No parcel-level resolution. No scenario-based projections. No risk score.
Traditional models weren't built for this.
AI makes it possible to do this at scale, at speed, at the resolution decisions actually require.
How the AI Works
Input 01
Satellite imagery
Landsat & Sentinel-2. Decades of coastal photographs from space, showing how shorelines have actually moved over time. The AI learns retreat trajectories directly from observation — not from formulas.
Input 02
Climate data
CMIP6 · ERA5. Global models forecasting sea-level rise, storm intensity, and ocean conditions under different emissions futures (SSP1-2.6 through SSP5-8.5). The AI conditions projections on scenario.
Input 03
Coastal physics
Wave energy, sediment dynamics, bathymetry, InSAR elevation data. How waves hit the shore, how sand moves, and the shape of the land that water meets. Physics constrains what the AI can predict.
Architecture
Deep learning model
Trained on multi-decadal satellite time series · attention across space and time · parcel-level resolution · multi-scenario output
Output
Parcel-level projections across 10, 30, 50, and 100-year horizons. Multiple erosion scenarios. One clear risk score.
No GIS expertise required. Output is designed to be read by elected officials and developers without specialist training.
The Output
Parcel-level projections across 10, 30, 50, and 100-year horizons. Multiple erosion scenarios. One clear risk score.
Soulac-sur-Mer · Source: CEREMA / IPCC AR6
What the AI Sees in the Data
Chronic erosion only — storm events excluded · Soulac-sur-Mer reference point
Sources: CEREMA · 4.3 m/yr average 1997–2021 · Région Nouvelle-Aquitaine · up to 8 m/yr (current accelerated rate)
Responsible AI
The danger isn't that AI is wrong. It's that it looks right.
AI outputs can look precise. Uncertainty must be visible, not hidden. AquAdapt treats confidence intervals as a primary output, not a footnote. Every projection ships with its full scenario envelope — so the range of futures is as legible as the forecast itself.
A model trained on the past cannot guarantee the future. AquAdapt does not extrapolate beyond what the data supports. Projections are conditioned on emission scenarios, not presented as deterministic outcomes. The model distinguishes between what it has learned and what it is inferring.
Knowing the limits of a model is part of operating it responsibly. AquAdapt is explicit about geographic scope, training data coverage, and confidence degradation over time. Where the model is less reliable, it says so — quantified, not qualified.
What the Output Looks Like
AquAdapt does not give a single answer. It gives the full picture — how much retreat to expect under each emissions pathway, and how certain the model is. This is what responsible AI looks like.
Parcel identifier
33-2841-B · Soulac-sur-Mer (Gironde)
Distance to shore today
147 m
SSP1-2.6
low emissions
SSP2-4.5
intermediate
SSP5-8.5
high emissions
10 yr
~2035
Safe
−18 m
[−12 · −24]
High conf.
−21 m
[−15 · −28]
High conf.
−26 m
[−19 · −34]
High conf.
30 yr
~2055
Safe
−38 m
[−28 · −50]
High conf.
−46 m
[−34 · −60]
High conf.
−62 m
[−48 · −80]
High conf.
50 yr
~2075
At risk
−58 m
[−44 · −74]
Mod. conf.
−82 m
[−63 · −104]
Mod. conf.
−118 m
[−94 · −148]
Mod. conf.
100 yr
~2125
Critical
−104 m
[−78 · −136]
Low conf.
−152 m
[−118 · −196]
Low conf.
−214 m
[−172 · −268]
Low conf.
Training data: 1984–2024
Intervals: 90% credible · Monte Carlo n=500
Confidence degrades at longer horizons — disclosed explicitly
Confidence intervals widen at longer horizons — not because the model is failing, but because climate uncertainty compounds over time. AquAdapt discloses this explicitly. A 100-year projection under SSP5 is not a prediction. It is a conditional bound.
AquAdapt Makes One Thing Visible
This is the question coastal local governments, real estate developers, and infrastructure developers cannot currently answer. AquAdapt answers it — at the resolution that decisions actually require, with the honesty that responsible AI demands.
Let's Talk
First pilots targeted Q2 2027. MVP delivery January 2027. Primary targets: the 371 local governments subject to the décret trait de côte, coastal real estate developers, and infrastructure developers operating in coastal zones.
Cité Universitaire · Paris 2026