Large Scale Space System Optimization
Business Context: Blue Origin will land humans on the Moon this decade. To achieve this, three spacecraft must cooperate as one system, requiring extreme performance optimization to achieve this technical challenge.
Problem Statement: Develop a data and simulation pipeline to minimize spacecraft mass while meeting crew safety requirements.
My Role: Conceived, built, and applied an end-to-end system optimization pipeline.
Key Contributions:
- Built a Python library to automate NASA simulation runtime, integrated with HPC and ML (Dask, scikit-learn, PyTorch).
- Led development of finite element models and system architectures for three spacecraft, with total complexity comparable to the International Space Station.
- Implemented a genetic algorithm for a non-differentiable problem; designed a multi-domain fitness function, tuned with Bayesian optimization and surrogate models.
- Built and administered an on-premise compute cluster to scale simulations.
- Directed cross-team collaborations (Structures, Thermal, Fluids & Propulsion, Avionics, Communications, Mission Design) to implement vehicle-wide mass savings.
My Impact: Delivered ~200 kg mass savings to Blue Moon MK2 Crew Lander vehicle by removing unnecessary material while ensuring crew safety.
Multi-Agent Reinforcement Learning (MARL) of Competitive Satellite “Tag”
Business context: Large satellite constellations require autonomous spacecraft guidance systems to avoid collisions. I posed myself the questions: What behaviors and system optimizations might evolve if you were trying to maximize the probability of collision in a game of tag? How would I build the MARL?
Problem statement: Maximize the probability of conjunction for a satellite constellation (blue team) and minimize probability of conjunction for another constellation (red team).
My role: I ideated and implemented this project in my free time.
My deliverable and it’s implementation:
- Multi-Agent RL Framework: Built PettingZoo-compliant MARL system with competitive satellite constellations, 94-element observations, 7 actions, and adversarial team training.
- Space Physics Engine: Modeled atmospheric drag, collision debris, and spatial grid collision detection with O(n log n) performance.
- Game Mechanics: Designed “Space Tag” pursuit/evasion gameplay with scoring (+100 tags, –15 clustering) and variants like King of the Hill.
Training Pipeline: Created curriculum learning (0–100 debris), adaptive team balancing, multi-objective optimization (competition, safety, fuel), and Proximal Policy Optimization (PPO) with reward shaping and clipping for stable multi-agent training
Software Architecture: Delivered modular Python package with physics components, test suite, CLI, and automated visualization in JS/React.
- Scalability & Performance: Optimized for 100+ satellites, enabling rapid training (80 episodes/minutes) with real-time analytics and behavior tracking.
My Impact: Developed a competitive Multi-Agent Reinforcement Learning framework for satellite constellations, training hundreds of agents in realistic space physics to demonstrate scalable coordination strategies for space traffic management, collision avoidance, and defense-tech.