Mining Saturn's Rings

View on GitHub → Updated 2026-07

Project ICEBERG is a mission-concept study: fly a tug to Saturn's B ring, trawl a 40-tonne block of nearly pure water ice into a fabric bag, and spend seven years flying it home while drinking it -- the cargo is the propellant tank. I ran a 167-round desk-study campaign against the concept, with pre-registered hypotheses and Monte Carlo sweeps over the full architecture space, trying to falsify it. The concept survived, in a narrow and heavily conditioned way: closure hinges on three specific engineering bets, each unproven, each mapped to a demonstrator objective. This is the writeup of the architecture, the method, the things the campaign killed, and the honest residue.

Abstract

The orbital economy runs on water -- propellant, radiation shielding, life support -- and today every kilogram of it is launched out of Earth's gravity well. Saturn's B ring is the largest deposit of nearly pure water ice in the solar system: 99.5 to 99.9 percent water by volume, pre-broken into house-sized and smaller blocks, sitting in a microgravity environment where a spacecraft can match orbits with a chunk at millimetres-per-second closing speed. ICEBERG asks what it would take to bring some home. The architecture is a water-fed microwave electrothermal tug with a kilowatt-class fission reactor and one genuinely new subsystem: a sublimation-capture trawl bag that turns the captured chunk into the return-leg propellant supply. A 167-round analysis campaign -- pre-registered hypotheses, runnable rounds, a Monte Carlo mission-graph framework sweeping 10 mission phases by roughly 36 options per phase -- produced a conditional verdict. At a 25-tonne commercial delivery floor, 5,656 feasible paths close across 322 unique architectures, best case 39.5 tonnes delivered in an 11.93-year round trip. All of them assume three things nobody has demonstrated: active capture of a multi-tonne ring particle, continuous months-long water-electrothermal operation on ring-sourced water at a specific impulse only reached in laboratory pulses so far, and a flight fission reactor delivered on program schedule. The campaign's job was to find those three bets and price them. The economics land in regulated-utility territory, and the study says so.

1. The premise

Water is the bulk commodity of any orbital economy that grows past boutique scale. It is propellant (electrolyzed or heated directly), radiation shielding, thermal mass, and the thing humans drink. Every kilogram of it in orbit today rode a rocket up a gravity well at launch prices measured in thousands of dollars per kilogram.

The standard answer to that problem is lunar polar ice -- and for the first decade of in-space water demand, the study concedes lunar in-situ resource utilization is the better architecture. The lunar deposits are close and reachable with known mission classes. They are also dilute: the LCROSS impactor measured 5.6 ± 2.9 percent water by mass in the best-case cold trap at Cabeus crater, and average regolith runs around 1 percent. Mining it means excavating, hauling, and thermally processing 20 to 100 tonnes of abrasive rock per tonne of water, at cryogenic temperatures, in craters that have never seen sunlight.

Saturn's B ring is the opposite deposit. Cassini microwave radiometry put the B ring at 99.5 to 99.9 percent water ice by volume -- a purer ore than anything ever mined on Earth -- and the main rings overall above 99 percent. The deposit is pre-broken: ring particles run from dust grains to metre-class boulders, orbiting in a sheet where relative velocities between neighbouring particles are millimetres per second. There is no digging. You match orbits and drift into the material. The ore grade is 18 to 100 times better than lunar polar, the extraction energy is nearly zero, and the catch -- the entire catch -- is that it sits 1.2 billion kilometres away and 13 years round trip.

ICEBERG is the study of whether that trade ever makes sense.

2. The architecture

One ship, ten phases, about 13 years:

1. Launch and kick. A Falcon Heavy lofts a roughly 50-tonne stack to low Earth orbit; a chemical kick stage performs the 7.3 kilometres-per-second trans-Saturn injection and is jettisoned. This is the only chemical propulsion in the entire mission.
2. Cruise and capture. A 6.1-year Hohmann transfer, then multi-pass low-thrust capture at Saturn on the water thruster -- the same low-thrust orbit-insertion approach NASA's Dawn mission used at Vesta and Ceres.
3. Trawl. The ship drops into a circular orbit at the B-ring radius and deploys a fabric trawl bag, lined with ballistic weave and aerogel. Induced radial drift at millimetres per second pushes ring material into the intake; soft layers decelerate particles inelastically. Light-time to Saturn is about 83 minutes, so the trawl runs on full autonomy -- the moment the program lives or dies, and nobody on Earth finds out for an hour and a half.
4. Cinch and convert. The bag closes. The sun-facing wall sublimates ice; the cold wall cryopumps the vapor back to frost; a heated harvest port meters vapor on demand into the propulsion feed. The cargo is simultaneously cargo, tank, and thermal management system.
5. The long ride home. Multi-pass Saturn departure, seven years of continuous chunk-fed thrust, then a multi-flyby lunar-gravity-assist capture at Earth -- flight-proven trajectory class (Hiten, WIND, Geotail, ARTEMIS) -- and a low-thrust spiral to the depot. Aerocapture is dead on arrival for this cargo: tens of tonnes of ice arriving at interplanetary speed would come through an atmosphere as vapor.

The propulsion is a water-fed microwave electrothermal thruster: a resonant cavity heats water vapor directly, no electrodes in the plume, no grids to erode, contamination-tolerant by construction. Laboratory thrusters of this class run in the 700-to-900-second specific-impulse band on water in short pulses, and a NASA small-business program targeted more than 800 seconds -- comparable to hydrogen-oxygen chemical rockets and about 2.5 times better than the best storable chemical propellants, from a propellant you can scoop out of the sky. Power comes from a kilowatt-class fission reactor in the Kilopower lineage. Both of those sentences carry asterisks big enough to get their own section below.

3. The bag

Everything else in the stack -- water thruster, small reactor, deep-space autonomy, lunar gravity assists -- exists somewhere on a technology-readiness ladder. Nothing like the sublimation-capture trawl bag has ever been built, and the study treats it as the central engineering risk. It also turns out to be the subsystem where the physics is friendliest.

The reason is the closing speed. Ring particles in neighbouring orbits drift past each other at millimetres per second, so capture happens at the kinetic energies of falling office supplies. A one-metre, 500-kilogram ice boulder entering the bag at 1 millimetre per second carries about 0.25 millijoules -- the energy of a paperclip dropped from an inch up. Even a 10-metre, 470-tonne boulder at the same closing rate carries only a quarter of a joule. Energy never threatens the fabric. The threat is geometry: a 10-metre boulder does not fit through the intake, so the design carries a mesh pre-screen sized to the largest admissible body and a forward-looking lidar that triggers a reject maneuver when something oversized wanders into the approach corridor.

The other two bench questions are aging and bookkeeping. The bag laminate -- ballistic weave lined with aerogel -- has to hold vapor loss below roughly 5 percent of cargo mass across a seven-year inbound coast at sun-facing wall temperatures of 200 to 280 kelvin; leak more than that and the chunk-fed delta-v budget stops closing the rocket equation on the way home. Whether current materials meet that number, or the liner needs metallization, is a bench-test answer. And as the bag fills, the stack's center of mass walks and the orbit perturbs; holding a few-hundred-metre stationkeeping box through a multi-hour fill costs on the order of 10 metres per second of thruster delta-v -- small, and carried in the budget.

Three named engineering questions, zero physics vetoes. That profile -- everything hard about the subsystem is testable on Earth for bench-test money -- is what lets the staged program structure in section 8 work at all.

4. The method: 167 attempts to kill it

A concept like this generates motivated reasoning the way a gravity well generates delta-v requirements. Anyone can write a pitch that sounds good. The interesting question is whether the concept survives an adversarial analysis campaign run under rules that make quiet self-deception expensive.

The campaign protocol is adapted from machine-learning experiment conventions and lives in the repo (PROTOCOL.md). Each round is a directory with a pre-registered hypothesis stated before the run -- falsifiable, with numeric ranges -- a runnable run.py, results committed unedited, and a mandatory Revisit clause: did the prediction hold, and if it was wrong, why. Findings sort into adopt, drop, or defer. Retractions stay in the record; several headline claims in the architecture decision matrix are marked RETRACTED in place by later audit rounds, with the audit trail pointing at what killed them.

The heavy machinery is a mission-graph framework: the full mission decomposed into 10 phases (launch, assembly, outbound, capture, harvest, departure, inbound, Earth arrival, and processing variants), roughly 36 options per phase, swept as a Monte Carlo over the physics anchors. A closure predicate -- delivered mass above a floor, round trip under a ceiling, every phase physically consistent -- filters the space. The canonical sweep evaluates tens of thousands of paths per run.

The work was carried out by AI research agents -- Claude, running as parallel orchestrated sessions, each session working assigned rounds against the shared protocol -- with a human owner setting direction, locking audited findings as ground truth, and issuing kill directives when a line of analysis got ahead of the evidence. Round documents attribute work to session codenames from Saturn's moons: titan, rhea, phoebe, hyperion. I am comfortable calling this a new way of doing desk studies; a 167-round pre-registered campaign is roughly a person-year of analyst work, and it ran in about ten weeks on evenings. The failure modes are new too, and section 9 is honest about them.

5. The graveyard

The campaign's most useful outputs are the things it killed. A sample, from the decision matrix's retraction trail:

The pitch document in the repo carries its own scars -- reader's notes flagging which tables did not survive later integration, left in place as audit trail rather than silently rewritten. That choice, made early, turned out to be the campaign's most valuable convention: you can trace every headline number backward to the round that set it, and every retired number to the round that killed it.

6. What survived

After the corrections, the canonical sweep at a 25-tonne commercial delivery floor finds 5,656 feasible paths across 322 unique architectures. The best single delivery is 39.5 tonnes in an 11.93-year round trip. The dominant closing pattern is almost boring: single launch, autonomous assembly, low-thrust spiral out, chunk-fed spiral home, direct propulsive or lunar-assisted arrival. The exotic options the framework also encoded -- Venus-Earth gravity assists, lunar-orbit processing waypoints, Saturn-moon flyby captures -- close nothing that the plain path does not close better. And the physics floor matters less than one programmatic number:

The closure verdict is a function of where the customer sets the floor more than of any single physics axis. A demonstrator-class mission closes handily. A 50-tonne commercial floor closes nothing at current anchors. The commercial case lives entirely in the 20-to-40-tonne band, and the requirement document now carries the floor as its most load-bearing open decision, set provisionally at 25 tonnes -- 5 tonnes below the engineering ceiling -- pending a real financial model.

The matrix above is the audit sweep rendered as the close/not-close dashboard view: 48 cells, and every ink-ringed closing cell sits in the corner that demands both flight-unproven bets at once. An interactive version -- pick your own delivery floor, hover any cell for its margin -- lives at robotrocketscience.com/projects/iceberg-matrix.

Two other survivals are worth naming. Small vehicles beat big ones: the 50-to-200-tonne vehicle grid the study started with was launcher-anchored habit, and 10-to-50-tonne vehicles close as well or better. And the trawl-bag physics -- the one new subsystem -- generated three bench-testable engineering questions and zero physics vetoes. Particle kinetic energies at ring closing speeds are millijoules; the problem is geometry and materials aging, which is what bench tests are for.

7. The three bets

Reduce the whole campaign and the verdict fits in three lines. Closure requires, jointly:

Bet 1: active chunk capture at scale. The desk-study anchor assumes 85 percent single-pass capture efficiency. A bottoms-up decomposition -- rendezvous times deployment times catch times containment times survival -- multiplies out to roughly 46 percent. The sweep says the matrix collapses below about 75 percent of the anchor: at half efficiency the best delivery drops below 20 tonnes and the commercial floor is gone. Cassini flew through ring material and sampled it passively; nobody has actively captured a multi-tonne object in a ring environment. This is the bet with no flight precedent at any scale.

Bet 2: continuous water electrothermal at flight scale, on ring water. The ground evidence is real: laboratory microwave electrothermal thrusters have reached the 700-to-900-second specific-impulse band on water vapor in short pulses, and Momentus has flown water-fed microwave electrothermal propulsion on its Vigoride vehicles. The gap is the operating profile. Lab results come in 50-second pulses on clean water; the mission needs continuous operation for months on B-ring water carrying about 0.3 percent non-icy contaminants through a resonant cavity. The sweep prices the sensitivity: the 800-second anchor sits at the closure cliff's inflection point -- 100 seconds lower halves the closure rate, 200 seconds lower collapses it.

Bet 3: a flight fission reactor, on schedule. The study's power section is the bleakest reading in the repo. The United States has orbited exactly one fission reactor -- SNAP-10A, in 1965 -- and every flight program since has died before orbit: SP-100, Timberwind, Prometheus, and most recently DRACO, cancelled in 2025. Six programs, roughly 1.7 billion dollars, zero orbits. The KRUSTY ground test proved the physics at 2.4 watts per kilogram system-level; a flight program has not been funded, and NASA's Fission Surface Power effort had not awarded Phase 2 as of the study's close. The campaign treats reactor availability as a Bayesian prior of program failure, applied on top of the physics, and hard-conditions the commercial case on a flight reactor being on contract before fleet money moves.

Each bet fails differently -- the first on mechanism, the second on endurance, the third on programmatics -- and each maps to a demonstrator objective that buys it down for two orders of magnitude less than the mission it de-risks.

8. The economics, stated plainly

The pitch's original framing was venture-scale returns from sovereign-scale cash flows. The campaign's financial rounds dismantled that framing and replaced it with something smaller and more defensible.

Run the program as a single upfront commitment and the mean expected net present value is negative -- the loss cases dominate the probability mass, and the upside lives in the top decile of water-price outcomes. Run it as staged options -- demonstrator gates every couple of years, kill criteria at each gate, fleet capital committed only after Saturn capture is proven -- and the numbers change shape rather than sign: the gate-0 commitment drops to about 11 percent of program capital, the option to abandon is worth about 2.3 billion dollars against the fleet decision, and the smart gate kills the fleet in 65 percent of Monte Carlo draws. Staged commitment is a downside-protection story, and the study says so in those words. What survives is infrastructure logic: if the demonstrators close the three bets, the delivered product prices like a regulated utility with a physical moat -- the second entrant is 13 years behind by orbital mechanics, no matter how much money they raise.

I want to be precise about what the study does and does not claim here. It does not claim ICEBERG is a good investment; the mean-EV rounds say the opposite at today's anchors. It claims the option structure is cheap relative to what it purchases, the downside is bounded at the first gate, and the price of finding out is three demonstrator missions of ordinary ambition.

9. What this method is worth

This project is also an experiment in how far AI-agent analysis campaigns can carry a mission concept, and I owe the reader an honest accounting of both directions.

What worked: pre-registration with mandatory revisits caught the campaign's own errors at a rate I have never achieved solo -- the delta-v anchor corrections, the retracted closure headlines, and the pitch-table falsifications all came from later rounds auditing earlier ones under standing rules. The mission-graph framework made "did you check the weird architecture" a query rather than an argument. Parallel sessions meant the audit rounds genuinely did not share context with the rounds they audited.

What the method cannot do: it cannot bend metal, and it inherits the quality of its anchors. Every closure number in this writeup is desk-study grade -- sourced to public literature, swept for sensitivity, and still a paper number. A bench test of bag-laminate permeability outranks every Monte Carlo in the repo, and the study's own conclusion is that the next dollar belongs in hardware questions, not more sweeps. There is also a subtler failure mode: agent campaigns generate analytical confidence faster than they generate evidence, and the volume of internally consistent documentation can read as maturity it has not earned. The kill directives in the audit trail exist because the owner watched for exactly that.

10. Conclusion

Somewhere in Saturn's B ring there is a house-sized block of water at ore grades no terrestrial mine has ever seen, and the physics of going to get it are mostly settled physics. A 167-round campaign could not kill the concept outright -- at a 25-tonne floor, thousands of architecture paths close on corrected anchors. What the campaign did instead was compress every open question into three bets: catch it, drink it, power it. Nobody has won any of the three. All three are demonstrable at demonstrator scale for a small fraction of program cost, and the staged structure means nobody has to bet the fleet to find out.

The idea stays filed under honest long shots. The method -- pre-registered adversarial desk studies at campaign scale -- is the part I expect to reuse next month.

Update 2026-07-20 -- the de-spin question

Rounds 168 and 169, run after this article first went up, chased an owner question: de-spinning a captured chunk has to cost propellant, and that propellant is delivered water -- how much?

The answer split in half. De-spin proper is a non-problem. Ring chunks rotate slowly -- the campaign's Cassini-anchored prior puts the median near 0.005 revolutions per minute, because collisions re-equilibrate ring-particle spin every orbit -- and at those rates removing the angular momentum of even a 200-tonne porous chunk costs grams to single kilograms (R_chunk_despin_budget). Three of the round's five pre-registered bounds still fell, all because my back-of-envelope used the solid-ice corner of a grid whose porous corner carries 1.5 times the moment of inertia; the numbers were wrong, the conclusion survived, and the lesson -- set bounds at the worst corner of your own grid -- is recorded in the study. The passive version is cheaper still: at any plausible cinch tension, the bag's own fabric friction de-spins the cargo into the stack in minutes. One bug also got caught on the way: the first closure reprocess read a results field that did not exist, scored every cell zero, and "passed" over an empty denominator -- a reminder that a hypothesis that holds over zero cases has not been tested.

The second half is where the owner's instinct pointed, one mechanism over (R_com_offset_thrust_alignment). The inbound burn drinks roughly 40 percent of the cargo through a harvest port. Drawn from one side, the cargo's center of mass walks 1.2 to 1.7 metres off the thrust line -- offset is the default trajectory of this architecture, a corner case only if the draw is actively symmetrized. Fight that torque with the corner reaction-control thrusters and the bill runs 50 to 500 tonnes of propellant, several deliveries' worth, at any power level: the angular impulse is fixed by the delta-v, so a bigger reactor does not help. Steer the thrust vector through the walking center of mass instead and the whole problem collapses to a cosine tax bounded at 2 percent of inbound propellant -- 31 to 492 kilograms on the canonical 40-tonne chunk. On the audit sweep the two strategies separate perfectly: steer-through flips zero of the five closing cells, and fighting it kills all five.

So the spin tax, quantified: grams for the de-spin, tens to hundreds of kilograms for steering through the walk -- and a new hard requirement for the vehicle, roughly 12 degrees of thrust-vector authority tracking a plus-or-minus 1.7 metre center-of-mass walk (or an actively symmetrized harvest draw). The ship specification carries neither today.

A third round (R_harvest_draw_symmetrization) then attacked that requirement's own premise -- the one-sided draw is a design choice, not a law. Verdict: put the harvest port on the thrust axis and roll the stack about it once a day, Apollo-style, and the walk collapses to centimetres; the 12-degree tracking requirement demotes to a no-roll fallback, replaced by a 3-degree trim plus a port-centering tolerance. Three of that round's four pre-registered bounds fell too, and one of the misses was the round's best finding: the walk amplification scales with the cargo's own draw fraction, so small chunks -- easy everywhere else in this campaign -- are the hard case here, needing the port centered to about 7 centimetres at 25 tonnes. The requirement now sitting with the campaign's decision record is the cheap version, with the expensive one as fallback.

Technical details

• Corpus: concept of operations, requirements (L0/L1) with variance log, risk register, 22 architecture-decision-record axes, 167 analysis rounds with pre-registered hypotheses and runnable rounds, Monte Carlo mission-graph framework (10 phases, ~36 options per phase).
• Stack: Python 3.13, numpy, scipy, Cantera for nozzle thermochemistry, matplotlib; uv-managed; physics models unit-tested.
• Method: pre-registered rounds per PROTOCOL.md; AI research agents (Claude) in parallel orchestrated sessions; human-locked ground-truth findings; retractions kept in the record.
• Verification: every headline number in this article traces to a round document in the repo; sensitivity sweeps cited inline; corrected anchors derived from primary sources (Edelbaum bounds, vis-viva, Cassini radiometry, LCROSS).
• Everything is public: the full campaign paper trail ships in the repository, including the retractions.

github.com/robotrocketscience/iceberg →