The point of this tool isn't precise prediction - it's to show how the pieces fit together. But the inputs are picked to match real research, not invented. Here's the grounding for each assumption, with sources, and an honest note on how solid each one is.
- NASA CP-2255, "Advanced Automation for Space Missions" (1980), Ch. 5 - the
original self-replicating lunar factory study (Freitas, von Tiesenhausen et al.).
- Coined the "vitamins" framing: build 90–95% of the factory locally and ship the last 5–10% (mostly electronics) from Earth.
- Estimated 90–96% mass closure is achievable; chasing 100% isn't worth it.
- Used a 1-year doubling time and showed exponential growth.
- Power: ~1.7 MW (solar) for its 100-tonne seed (design range 0.47–11.5 MW).
- https://www.rfreitas.com/Astro/GrowingLunarFactory1981.htm · https://ntrs.nasa.gov/citations/19830007081
- Freitas & Merkle, "Kinematic Self-Replicating Machines" (2004) - the modern synthesis. Defines matter / energy / information closure and the key dynamic this tool models: below full closure you're stuck waiting on resupply; only near 100% closure does growth become self-sustaining and exponential.
- M. V. Shubov, "Guided Self-Replicating Factory for Colonization of Solar System" (arXiv 2110.15198, 2021) - recent revival. Doubling time <1 year → large colony in ~two decades; near-term realistic closure ~70%; chips/solar/circuitry stay Earth-sourced.
Honesty note: there is no validated modern end-to-end seed-factory design. The 1980 NASA study is still the canonical reference. Treat every number here as "right order of magnitude," not a forecast.
| Assumption (this tool) | Real-world grounding | How solid |
|---|---|---|
| Seed mass 7–12 t | NASA strawman was 100 t; modern thinking favors smaller seeds | A modeling choice. Lighter than canonical - labeled as "modern/optimistic." |
| Power ~2–4 MW | NASA nominal 1.7 MW (range 0.47–11.5 MW), solar | Matches the canon. But a multi-MW plant outmasses the seed - so power is treated as separate infrastructure, not part of the seed. |
| Build rate 15–20 kg/day | NASA implies ~27 kg/day at equivalent productivity | Reasonable, slightly conservative. |
| Doubling ~1 year | NASA & the 2021 paper both use ~1 year | Best-supported number here. |
NASA Fission Surface Power (the real near-term lunar reactor) is 40 kW under 6 t: https://www.nasa.gov/centers-and-facilities/glenn/nasas-fission-surface-power-project-energizes-lunar-exploration/
These are the electricity the factory spends on-site to make 1 kg of a part. The headline fact - chips cost thousands of kWh/kg, smelted metal costs single digits - is what makes the electronics wall real.
| Part type (this tool) | Value used | Real range (sources below) | Verdict |
|---|---|---|---|
| Smelted/cast metal structure | 5 | 1.7–9.7 (recycled→primary steel) | OK for in-situ electric smelting |
| Thermal radiators | 3 | 1.7–5.6 (recycled metal) | OK |
| Refining-plant structure | 7 | 6–24 (steel→stainless) | Low-conservative |
| Actuators / motors | 15 | ~14–20 | Good |
| Robotic manipulators | 18 | similar to motors | Reasonable |
| Precision bearings / alloys | 35 | 20–46 (alloy/superalloy) | Good |
| Machined sensor housings | 60 | tens (small precision metal) | OK for metal only |
| Solar arrays (silicon) | 50 | 40–120 | Raised from a too-low 8; silicon purification dominates |
| Power electronics / ICs | 2,500 | 1,000–3,000 | Good |
| Electronic sensors | 4,000 | 2,000–8,000 | Raised from a too-low 600 |
| Compute / logic chips | 8,000 | 3,000–15,000 (finished, packaged) | Raised from a too-low 2,000 |
The chip number deserves a caveat. Embodied energy of silicon swings enormously with how you measure it: ~1,800 kWh/kg for a blank wafer, ~3,000–15,000 kWh/kg for a finished packaged chip, and 100,000+ kWh/kg if you count only the active silicon die. We use a finished-packaged-chip basis (8,000). On any basis, chips are the most energy-expensive thing in the factory by a wide margin.
Paper figures (src/closure_sim/paper_figures.py, for papers/electronics-wall/).
The three figures in the electronics-wall paper restate this module's deterministic output
and pure derivations; no new numbers. fig_leverage.pdf plots 1/(1-C) (mass balance),
marked at C=0.67 -> 3.0x and C=0.97 -> 33.3x. fig_embodied_energy.pdf plots the
per-subsystem energy_to_produce_kwh_per_kg values in the table above, read from the
loaded factory. fig_chip_crossover.pdf sweeps electronics_wall(...) over available
power for the lunar seed (closure 97.08%): importing chips is resupply-limited at ~28.8 yr
(10,512 d), making them locally is energy-limited at ~17.4 yr (6,350 d) at 4 MW and never
completes near 1 MW. Regenerate via uv run --extra dev python -m closure_sim.paper_figures
(days converted at 365.25 d/yr).
Sources:
- Williams, Ayres & Heller, "The 1.7 Kilogram Microchip", Env. Sci. Technol. 2002 - inputs to make a 2 g chip outweigh it ~600×. https://pubs.acs.org/doi/10.1021/es025643o
- Nagapurkar & Das (Oak Ridge NL), IC manufacturing energy, Sust. Mat. & Tech. 2022 - 9–38 MJ/cm² of wafer. https://www.osti.gov/servlets/purl/1884036
- Metals/solar embodied energy: Inventory of Carbon & Energy (ICE) coefficients https://www.wgtn.ac.nz/architecture/centres/cbpr/resources/pdfs/ee-coefficients.pdf ; Peng et al. 2013 PV life-cycle review https://krichlab.ca/wp-content/uploads/2014/06/Peng2013_Review-LCA-EPBTGHG-SolarPV.pdf ; Marspedia embodied energy https://marspedia.org/Embodied_energy
- Power-electronics LCA (Si IGBT vs SiC): https://www.sciencedirect.com/science/article/pii/S2772370423000184
- Pirson & Bol, "Assessing the embodied carbon footprint of IoT edge devices with a bottom-up life-cycle approach" (arXiv 2105.02082, 2021) - sensor/edge-device embodied footprint: https://arxiv.org/abs/2105.02082
- Why the chip supply chain is the deepest on Earth (400+ steps, 9N-pure materials, EUV monopoly): CSIS https://www.csis.org/analysis/mapping-semiconductor-supply-chain-critical-role-indo-pacific-region
Every entry in the per-part energy table now has a companion Uniform(low, high)
distribution in src/closure_sim/distributions.py (PART_ENERGY_KWH_PER_KG_DIST),
using the LCA min-max endpoints exactly as reported in the table above. The
sintered-regolith strength range - explicitly labelled here as "carry as a band,
never a point" (>100x span across techniques) - lands as LogUniform(2.49, 355.0)
so each order of magnitude is equally likely, matching how the source presents the
choice of technique rather than the linear numeric distance.
A first UQ finding surfaced by MC + Sobol over the per-part bands: for the ratio leverage = chip_kWh_per_kg / metal_kWh_per_kg (the electronics-wall driver), metal energy dominates the Sobol total-order ranking, not chip energy - the small denominator amplifies proportionally-small changes into larger changes in the ratio than the wide chip numerator does. This is exactly the "which input actually drives this finding" attribution issue #35 asks the papers to report.
ROADMAP-PROPOSAL.md weighed making structures its own module vs a parameter here. The
default was to demote it to a mass-penalty parameter unless the k=1.0 regression showed
the physics moved real closure numbers. It was demoted; structures.py is that parameter.
- Sintered lunar-regolith compressive strength: ~2.49-355 MPa, a >100x span by
technique.
SINTERED_REGOLITH_STRENGTH_MPA(representative points): solar-3D/PBF ~4.2 MPa, microwave-sintered ~37 MPa (KLS-1), sintered ~98 MPa (air) / ~152 MPa (vacuum), dense traditional ~232 MPa, glass-ceramic (800 C) ~355 MPa. Sources: "Solar 3D printing of lunar regolith", Acta Astronautica 152:800 (2018), https://www.sciencedirect.com/science/article/pii/S0094576518303874 ; microwave sintering of KLS-1, https://www.researchgate.net/publication/351670581 ; glass-ceramic from regolith simulant, https://www.sciencedirect.com/science/article/abs/pii/S0022309326000517 . Verdict: sourced; carried as a band, never a point (the >100x span is the whole reason strength can't be a single number). - Mass penalty
k = required_strength / material_strength(clamped >= 1): a weaker material needs a heavier part for the same load. Derived, shown.
- k = 1.0 reproduces closure exactly (regression test): with no penalty,
closure_with_structural_penaltyreturnscompute_closure's ratio bit-for-bit. - A mass penalty raises closure, it does not lower it. Heavier local structure is still local mass against fixed imports, so weak material costs throughput and energy (more mass to build), not closure. A realistic k (microwave regolith, ~1.08) moves closure <1 point.
- Only a hard strength threshold moves closure - a part that cannot meet a
non-scalable requirement flips to an import (vitamin), which closure-sim already models
via
producible_locally. So the strength physics needs no new module: the mass penalty is a parameter, and the threshold is the existing boolean. That is the recorded verdict.
- Power is flat (doesn't grow as the factory grows) - a v1 simplification that makes the "energy wall" visible. A later module will let power scale.
- One number per part type for manufacturing energy - real LCA depends on which metal, recycled vs virgin, which chip node, etc. The README spells out the recycled-metal and packaged-chip assumptions baked in.
- Terrestrial energy figures are used as stand-ins; an off-world factory running on its own electricity would re-base some of these. Doesn't change the conclusion (chips ≫ metal).
Sources that ground this module's ideas or cross-check its numbers, consolidated in the project bibliography (frontend/src/sources.ts) and shown on the site's Sources page. These add context; they are not new numbers in the code.
- Metzger et al. 2013 - P. T. Metzger, A. Muscatello, R. P. Mueller & J. Mantovani (2013). Affordable, Rapid Bootstrapping of the Space Industry and Solar System Civilization (arXiv:1612.03238). Journal of Aerospace Engineering 26(1):18-29, DOI 10.1061/(ASCE)AS.1943-5525.0000236. https://arxiv.org/abs/1612.03238. The modern quantitative counterpart to NASA CP-2255: ~12 t of landed hardware bootstrapping to 156-40,000 t of industrial assets over ~20 years via robotics and additive manufacturing, starting sub-replicating (teleoperated, importing vitamins) and spiralling toward autonomy, with electronics staying Earth-sourced. Grounds the seed-mass, doubling-time, and partial-closure-then-grow dynamics.
- Jones et al. 2011 (RepRap) - R. Jones, P. Haufe, E. Sells, P. Iravani, V. Olliver, C. Palmer & A. Bowyer (2011). RepRap - the Replicating Rapid Prototyper. Robotica 29(1):177-191, DOI 10.1017/S026357471000069X. https://www.cambridge.org/core/journals/robotica/article/reprap-the-replicating-rapid-prototyper/5979FD7B0C066CBCE43EEAD869E871AA. The best real-world data point on partial self-replication: an open-source 3D printer that prints a large fraction of its own parts but not motors, electronics, or rods, with measured reproductive spread. A terrestrial echo of the electronics wall - a machine can close on structure but must import the high-tech vitamins.
- Boyd 2012 - S. B. Boyd (2012). Life-Cycle Assessment of Semiconductors. Springer (from the 2009 Stanford PhD dissertation), DOI 10.1007/978-1-4419-9988-7. https://escholarship.org/uc/item/8bv2s63d. The most complete transparent process-level LCA of CMOS logic, DRAM, and flash across seven technology generations - the strongest independent anchor for the finished-chip embodied-energy figure behind the 8,000 kWh/kg headline, and how it moves with node and yield.
- Gutowski et al. 2009 - T. G. Gutowski, M. S. Branham, J. B. Dahmus, A. J. Jones & D. P. Sekulic (2009). Thermodynamic Analysis of Resources Used in Manufacturing Processes. Environmental Science & Technology 43(5):1584-1590, DOI 10.1021/es8016655. https://doi.org/10.1021/es8016655. Across 20 processes, electricity used per kg of material rises by orders of magnitude from conventional metal shaping (casting, machining) to vapor-phase semiconductor processes - the exergy-based, physics-grounded basis for the central claim that chips cost roughly 1,000x more energy per kg than smelted metal.
- Murphy et al. 2003 - C. F. Murphy, G. A. Kenig, D. T. Allen, J.-P. Laurent & D. E. Dyer (2003). Development of Parametric Material, Energy, and Emission Inventories for Wafer Fabrication in the Semiconductor Industry. Environmental Science & Technology 37(23):5373-5382, DOI 10.1021/es034434g. https://doi.org/10.1021/es034434g. A bottom-up per-wafer energy and materials inventory for the fab itself - grounds the blank-wafer basis end of the chip energy range (about 1,800 kWh/kg for a bare wafer vs. thousands for a packaged part), and documents why the measurement basis you pick swings the number.
- Ashby 2012 - M. F. Ashby (2012). Materials and the Environment: Eco-informed Material Choice (2nd ed.). Butterworth-Heinemann / Elsevier, ISBN 978-0-12-385971-6. https://shop.elsevier.com/books/materials-and-the-environment/ashby/978-0-12-385971-6. Standard-reference embodied-energy and carbon datasheets for common materials - grounds the cheap-to-make structural end of the per-part table (metals at single-digit to tens of kWh/kg) and is an independent cross-check on the ICE coefficients already cited.
- Guerrero-Gonzalez & Zabel 2023 - F. J. Guerrero-Gonzalez & P. Zabel (2023). System analysis of an ISRU production plant: Extraction of metals and oxygen from lunar regolith. Acta Astronautica 203:187-201, DOI 10.1016/j.actaastro.2022.11.050. https://ui.adsabs.harvard.edu/abs/2023AcAau.203..187G/abstract. Detailed off-world energy and hardware budgets for molten regolith electrolysis and FFC-Cambridge processing (e.g. a ~6,776 kg plant making 25 t/yr ferrosilicon plus oxygen). Directly addresses the open caveat that the project uses terrestrial smelting energy as a stand-in - this gives the actual in-situ kWh/kg for making structural metal on the Moon.
- von Neumann & Burks 1966 - J. von Neumann; ed. A. W. Burks (1966). Theory of Self-Reproducing Automata. University of Illinois Press, Urbana. https://archive.org/details/theoryofselfrepr00vonn_0. The origin of the idea: von Neumann's universal constructor proved a machine can build a copy of itself if it carries both a construction description and a way to copy that description - the information-closure half of the matter / energy / information framing every seed-factory claim rests on.
- Chirikjian 2004 (NIAC) - G. S. Chirikjian (2004). An Architecture for Self-Replicating Lunar Factories. NASA Institute for Advanced Concepts (NIAC) Phase I Final Report, study 880. https://www.niac.usra.edu/files/studies/final_report/880Chirikjian.pdf. A concrete subsystem architecture for exactly the object closure-sim models: a lunar factory that mines regolith, refines materials, and assembles copies of itself, decomposed into robots, refining, parts fabrication, and assembly. Grounds the what-must-a-real-factory-be-made-of breakdown and the teleoperation-to-autonomy path.
- Moses & Chirikjian 2020 - M. S. Moses & G. S. Chirikjian (2020). Robotic Self-Replication. Annual Review of Control, Robotics, and Autonomous Systems 3:1-24, DOI 10.1146/annurev-control-071819-010010. https://www.annualreviews.org/content/journals/10.1146/annurev-control-071819-010010. The modern survey tying the strands together: the principles required to make self-replicating robots from raw materials, the role of 3D printing, and the key distinction between closure of parts and closure of the fabrication processes that make them.
- Sagan & Newman 1983 - C. Sagan & W. I. Newman (1983). The Solipsist Approach to Extraterrestrial Intelligence. Quarterly Journal of the Royal Astronomical Society 24:113-121. https://ui.adsabs.harvard.edu/abs/1983QJRAS..24..113S/abstract. The canonical rebuttal to Tipler: self-replicating probes are inherently dangerous and hard to control (an unchecked replicator would consume the galaxy's mass), so a civilization would avoid or destroy them. Grounds the control / containment concerns and the sensitivity of outcomes to replication rate.
- Zykov et al. 2005 - V. Zykov, E. Mytilinaios, B. Adams & H. Lipson (2005). Robotics: Self-Reproducing Machines. Nature 435(7039):163-164, DOI 10.1038/435163a. https://www.nature.com/articles/435163a. A physical demonstration that mechanical self-reproduction is real, not just theory: modular molecube robots that pick up identical cubes from feeding stations and assemble a working copy. Grounds the plausibility of physical (not just computational) self-replication and the identical-modules-plus-feedstock model of closure.
- Kuehr & Williams 2003 - R. Kuehr & E. Williams (eds.) (2003). Computers and the Environment: Understanding and Managing their Impacts. Kluwer / Springer, Eco-Efficiency in Industry and Science vol. 14, DOI 10.1007/978-94-010-0033-8. https://doi.org/10.1007/978-94-010-0033-8. Established the material-intensity-of-computing case (a desktop PC takes on the order of 240 kg of fossil fuel, 22 kg of chemicals, and 1,500 kg of water to make) - the broader-context companion to Williams' 1.7 Kilogram Microchip, reinforcing why electronics are the hard-to-close vitamins.
- Elvis 2014 - M. Elvis (2014). How Many Ore-Bearing Asteroids? (arXiv:1312.4450). Planetary and Space Science 91:20-26. https://arxiv.org/abs/1312.4450. The scarcity of minable feedstock: only about 1 in 2000 accessible near-Earth asteroids is platinum-group ore-bearing and ~1 in 1100 is water-ore-bearing. Grounds the resource-availability assumption behind a probe or seed factory that must harvest local material rather than assume any asteroid will do.
- DeMeo & Carry 2014 - F. E. DeMeo & B. Carry (2014). Solar System evolution from compositional mapping of the asteroid belt. Nature 505:629-634, DOI 10.1038/nature12908. https://doi.org/10.1038/nature12908. The composition of the asteroid feedstock by taxonomic class and how it is distributed by mass and heliocentric distance. Grounds what a resource-harvesting module can expect to find where - which raw materials are actually available at a given mining destination.
- Lunar Sourcebook - G. H. Heiken, D. T. Vaniman & B. M. French (eds.) (1991). Lunar Sourcebook: A User's Guide to the Moon. Cambridge University Press (full text hosted by LPI/USRA). https://www.lpi.usra.edu/publications/books/lunar_sourcebook/. The definitive reference on lunar regolith and rock composition, mineralogy, and physical properties - the feedstock inventory for a factory that lands on the Moon and builds from local material. Grounds which elements (O, Si, Al, Fe, Ti, Mg) are locally available and in what abundance.
- Hoffman et al. 2022 (MOXIE) - J. A. Hoffman, M. H. Hecht, D. Rapp et al. (2022). Mars Oxygen ISRU Experiment (MOXIE) - Preparing for human Mars exploration. Science Advances 8(35), eabp8636, DOI 10.1126/sciadv.abp8636. https://www.science.org/doi/10.1126/sciadv.abp8636. The first demonstrated in-situ resource utilization on another planet: solid-oxide electrolysis of Martian CO2 producing ~6 g O2/hr on Perseverance. Grounds the Mars-destination ISRU case - a probe can make consumables from the local atmosphere rather than importing them.
replication.simulate delegates integration to vn_core.ode and records SimSteps at
the reporting cadence. Invariants live on the whole trace and are checked once by
_verify_trace_invariants(steps, alpha, energy_cap, seed_mass_kg), called under
if __debug__: at the end of simulate. python -O strips it. Positive + negative
tests live in tests/test_invariants.py.
- [inv:cs-mass-monotone] For every consecutive pair,
mass[i+1] >= mass[i] * (1 - 1e-8). Mass never decreases beyond RK45'srtol=1e-8; growth is one-directional in every regime currently modelled. - [inv:cs-mass-nonneg]
mass[i] >= seed_mass_kgfor all i (with1e-8slack). The seed is a floor. - [inv:cs-installed]
installed_capacity == alpha * factory_mass_kg(exact within float tolerance). This is a definition, not physics; asserting it catches bookkeeping drift. - [inv:cs-output]
output == min(installed_capacity, energy_cap). Same character as installed - a definition worth pinning. - [inv:cs-growth-nonneg]
growth_rate_kg_per_day >= 0. If any regime we later add introduces decay, this becomes conditional and is documented at that point.