Skip to content

troyrocket/space-machine

Repository files navigation

SPACE MACHINE

Multi-Domain AI Infrastructure Site Selection Engine — LEO · Moon · Mars

When Earth runs out of power and land, where does the next rack go?

Space Machine evaluates optimal locations for AI data centers across three domains using physics-based models, real NASA data, and interactive 3D visualization.


Three Domains

Domain Sites Altitude Latency Launch $/kg
🛰️ LEO 6 orbital configs 400–1,200 km 1–10 ms ~$100
🌕 Moon 8 surface sites 1.28 sec ~$400
🔴 Mars 8 surface sites 3–22 min ~$800

7 Evaluation Dimensions

# Dimension Formula Data Source
1 Solar Power solar_flux × availability × 100 NASA LRO/LOLA, Mars MOLA
2 Cooling εσA(T⁴_hot − T⁴_cold) + convective LRO Diviner, Mars climate
3 Communication earth_comm_availability × 100 Geometric + relay model
4 Terrain (1 − surface_difficulty) × 100 LOLA DEM, Mars HiRISE
5 Resources resource_proximity × 100 LCROSS, MRO radar
6 Radiation (1 − exposure) × 100 LRO CRaTER, Mars RAD
7 Cost (TCO) 9-component 10-year model Starship economics

Final Score = Σ(score_i × weight_i) / Σ(weight_i) — weights vary by workload type.

Quick Start

# Run full analysis (all domains)
python run.py

# LEO only
python run.py --domain leo

# Moon with batch training weights
python run.py --domain moon --type batch_training

# Mars with $5B budget
python run.py --domain mars --budget 5e9

# Terminal output only (no 3D visualization)
python run.py --no-viz

3D Visualization

The interactive visualization uses CesiumJS with real NASA terrain data:

  • Earth/LEO — Bing Maps satellite imagery + orbital ring polylines
  • Moon — NASA LRO Moon Terrain (asset 2684829, 1m polar resolution)
  • Mars — NASA MOLA + Viking colorized imagery (asset 3644333)

Setup

  1. Get a free Cesium Ion token at ion.cesium.com/signup
  2. Add Moon Terrain and Cesium Mars from the Asset Depot
  3. Run with token:
python run.py --cesium-token YOUR_TOKEN
# or
export CESIUM_TOKEN=YOUR_TOKEN
python run.py

Note: The visualization must be served via HTTP (not file://) for Cesium workers to load. run.py starts a local server automatically.

Top Results (Default: 50 PB Archival)

🥇 Starlink-Adjacent Orbit  [🛰️ LEO]   78/100   $261M
🥈 Equatorial LEO            [🛰️ LEO]   76/100   $259M
🥉 ISS Orbit                 [🛰️ LEO]   74/100   $261M
4. Connecting Ridge           [🌙 MOON]  74/100   $330M
5. Shackleton Crater Rim      [🌙 MOON]  73/100   $337M

Cost Model

10-year TCO with 9 cost components per site:

Component LEO Moon Mars
Launch cost/kg $100 $400 $800
Payload capacity 150,000 kg 50,000 kg 100,000 kg
Operations/yr $8M $12M $20M
Rad-hardening 2.5–5× 2.5–5× 2.5–5×
Cooling Radiative (free) Radiative (free) Radiative + CO₂ convective

Project Structure

space-machine/
├── run.py                     # CLI entry point
├── spacemachine/
│   ├── models.py              # SpaceSite, Workload, SiteScore
│   ├── sites.py               # 22 candidate sites (LEO/Moon/Mars)
│   ├── physics.py             # Domain-specific physics & cost models
│   ├── scoring.py             # 7-dimension weighted scoring engine
│   └── visualize.py           # CesiumJS 3D HTML generator
├── generate_ppt.py            # Goldman Sachs-style PPT generator
├── Space_Machine_Demo.pptx    # Presentation deck
└── space_machine_analysis.html # Generated 3D visualization

Physics Constants

Constant Value Source
Solar constant 1,361 W/m² NASA Solar Irradiance
Stefan-Boltzmann 5.67×10⁻⁸ W/(m²·K⁴) NIST CODATA 2018
Mars solar fraction 43.1% of Earth 1.524 AU distance
Lunar PSR temp ~40 K LRO Diviner
Laser comm bandwidth 1.2 Gbps (Moon), 0.1 Gbps (Mars) NASA LCRD
GCR dose ~0.7 mSv/day LRO CRaTER

Built With

  • Python 3
  • CesiumJS — 3D globe visualization
  • NASA LRO/LOLA, MOLA, Viking, MRO mission data
  • python-pptx — Presentation generation

Built at SemiAnalysis x FluidStack Hackathon, March 2026.

"You solved Dirt to Decode on Earth. We extended it to the stars."

About

Multi-Domain AI Infrastructure Site Selection Engine — LEO, Moon, Mars

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors