Research report

Satellite constellations for disaster response

Built and validated against simulation. No orbit has been flown.

Code

The question

Can a learned model stand in for hours-long orbit simulations, and find orbits that revisit disaster areas sooner?

43.9%

shorter mean revisit gap than a 48-hour standard benchmark, best simulated orbits across eight regions

The best simulated orbits cut the mean revisit gap by 43.9% against a 48-hour standard benchmark.

Standard benchmark
Simulator-backed data
Includes generated rows

Tap a region

Best simulated gap in each region's data, in hours, against a fixed 48-hour figure for one satellite on a standard sun-synchronous orbit. The benchmark is a textbook number written into the validation script, not a simulation. Mean of the eight regions: 26.9 h. Regions marked as including generated rows have a best gap that may come from a generated row.

Gaps as a table

Region	Best simulated gap (h)	Shorter than benchmark
Standard benchmark	48	-
Mean of eight regions	26.9	43.9%
United Kingdom	14.4	70.0%
Amazon	17.2	64.2%
Caribbean	19.9	58.5%
California	22.8	52.5%
Sahel	26.7	44.4%
Bay of Bengal	35	27.1%
Australia	38.7	19.4%
Pacific Northwest	40.7	15.2%

Why it matters

After a disaster, the gap between satellite passes sets how soon responders can see what changed. Searching orbits with a simulator is too slow to explore widely.

Method

A hybrid surrogate model is trained on 7,000 high-fidelity STK simulations to predict the average gap between observations over a region.
A multi-objective search (NSGA-II) looks for orbits against the surrogate, with a confidence bound on each prediction.
As a check, 500 orbits are compared against the simulator and against an independent analytical tool.

Data

11,200 rows over eight disaster regions: 7,000 from the simulator and 4,200 generated for the Sahel, Bay of Bengal and Pacific Northwest.
Held-out accuracy of the surrogate is R² 0.892 with a mean error of 13.65 hours; 98.7% of results pass the physics checks.

Results

A separate comparison: four satellites on optimised orbits against four on sun-synchronous orbits, about 43.6% shorter, the same factor in every region.

A four-satellite constellation on optimised orbits revisits every region 1.77× sooner than on sun-synchronous orbits.

Sun-synchronous
Optimised

Tap a region

Average gap between observations, in hours, for four satellites, from the project's constellation comparison. The factor is identical in every region, which the file does not explain; treat it as a property of the comparison, not of each region.

Constellation as a table

Region	Sun-synchronous (h)	Optimised (h)	Factor
Amazon Rainforest	9.2	5.2	1.772
Caribbean	9.8	5.5	1.772
Australia (NSW)	20.1	11.4	1.772
California	11.3	6.4	1.772
United Kingdom	7.1	4	1.772

The surrogate predicts shorter gaps in seven of eight regions, by up to about 32%, and a slightly longer one in the Sahel.

The surrogate predicts shorter revisit gaps in seven of eight regions, and a longer one in the Sahel.

Simulator-backed data
Includes generated rows

Tap a region

Improvement = (best simulated gap in that region's data - model-predicted gap at the optimised orbit) / best simulated gap. These are model predictions from the optimiser, not simulator runs of the optimised orbits, and the model reports LOW confidence in every region.

Regions as a table

Region	Best simulated gap (h)	Model-predicted gap (h)	Improvement	Surrogate R²	Upper bound (h)	Confidence
California	22.8	15.6	+31.7%	0.7134	90.8	LOW
Caribbean	19.9	13.6	+31.5%	0.8401	99.4	LOW
Amazon	17.2	12.1	+30.0%	0.8414	66.6	LOW
United Kingdom	14.4	11.4	+20.8%	0.707	107.8	LOW
Australia	38.7	35.4	+8.7%	0.8502	91.1	LOW
Pacific Northwest	40.7	37.5	+7.9%	0.9338	79.2	LOW
Bay of Bengal	35	34	+2.8%	0.9461	73.5	LOW
Sahel	26.7	27.9	−4.2%	0.9652	28.1	LOW

On 500 orbits the model tracks the simulator closely; the independent analytical tool agrees with it far less.

The model tracks the simulator closely (correlation 0.988); an independent analytical tool does not (0.45).

Tap a panel for its summary

500 orbits, 100 in each of 5 regions: gap in hours from the simulator (x) against the model and against the independent analytical tool (y). The dashed line is perfect agreement. The sample was drawn from the full dataset, so some orbits may also have been used in training.

Per-region summary as a table

Region	Orbits	Model correlation	Analytical correlation
Caribbean	100	0.991	0.503
Amazon	100	0.996	0.337
Australia	100	0.984	0.299
United Kingdom	100	0.985	0.624
California	100	0.980	0.377

Where the training rows come from: five regions are simulator-backed, three include generated rows.

7,000 rows come from the orbit simulator and 4,200 are generated.

Simulator rows
Generated rows

Tap a region

11,200 rows in all. Rows for the Sahel, Bay of Bengal and Pacific Northwest are generated, so results there describe the model, not the simulator.

Rows as a table

Region	Rows	Source
Australia	2,500	Simulator
Amazon	2,000	Simulator
California	1,500	Simulator
Sahel	1,400	Generated
Bay of Bengal	1,400	Generated
Pacific Northwest	1,400	Generated
Caribbean	500	Simulator
United Kingdom	500	Simulator

The honest limits

The 48-hour benchmark is a fixed figure for one satellite on a standard orbit, written into the validation script. It was not simulated here, and the best gaps for the Sahel, Bay of Bengal and Pacific Northwest may come from generated rows.
The per-region improvements are the model's predictions at the orbits it chose, not simulator runs of those orbits.
The model reports low confidence in every region. Its upper bound on the predicted gap is far above the prediction (for California, 90.8 hours against 15.6).
Three regions are trained partly on generated rows, so results there describe the model more than the simulator.
The 500-orbit check was drawn from the full dataset, so some orbits may also have been in training.
The four-satellite factor is identical in every region (1.772), which the results file does not explain.
No orbit has been flown. The surrogate is validated against simulation output, not on-orbit measurements.

What is next

Run the optimised orbits back through the simulator, to turn predictions into measurements.
Replace the generated rows for the three regions with simulated ones.

Next reportGlass Societies