AstroAI Workshop 2026
Derick F. Tangap
Cross-Scale Parameter Inference using Reinforcement Learning within a Centripetal Reference Framework
Presenter: Derick F. Tangap (Nebula Research Lab)
Title: Cross-Scale Parameter Inference using Reinforcement Learning within a Centripetal Reference Framework
Date/Time: Tuesday, June 16, 11:30 AM - 12:30 PM
Abstract: Abstract: Cross-Scale Parameter Inference using Reinforcement Learning within a Centripetal Reference Framework
Objective: This study proposes a unified computational framework for analyzing cosmic phenomena by bridging the gap between artificial inertial gravity and natural gravitational fields. We introduce a “Rotational Velocity Heuristic” that translates traditional Parameter Inference into a benchmarking system based on Earth’s kinematic constants. This allows Reinforcement Learning (RL) agents to more intuitively model the dynamics of Exoplanetary Systems, Gravitational Waves, and Large-scale Structures. Methodology: We treat gravitational environments as dynamic velocity-radius constraints. The RL agents optimize detection by utilizing Earth-Standard Units (ESU) as physical priors: Rotational Gravity Unit (1 $G_{eff}$): Modeled as the centripetal acceleration ($a_c = v^2 / r$) required to simulate $9.8 \, \text{m/s}^2$. In a spin-station context, the agent uses the tangential velocity relative to the hull radius to infer “weight.”
Orbital Velocity Unit (1 $V_{orb}$): Defined by Earth’s mean orbital speed ($\approx 29.8 \, \text{km/s}$), used as a reference point for calculating the gravitational potential wells of stellar bodies. The agents apply Time-series Analysis to transit and wave data, interpreting signal frequency shifts as variations in these fundamental velocity units. This “Physics-based Deep Learning” approach ensures that the agents remain bounded by the laws of angular momentum and General Relativity.
Results and Significance: By grounding RL agents in ESU benchmarks, the model achieves higher accuracy in characterizing high-mass environments (up to 20G equivalent) and detecting low-amplitude Gravitational Waves. The framework effectively filters stochastic noise by rejecting signal parameters that violate the $v^2/r$ relationship. This provides a scalable methodology for autonomous observatories to map the cosmic web and analyze exoplanetary atmospheres with increased physical fidelity.
Physical Foundations & Key Contributions Kinematic Benchmarking: Replaces abstract mass values with Earth-Standard Units based on tangible rotational and orbital velocities, simplifying the training of neural networks.
Centripetal Consistency: Ensures that “Spin Station” mechanics are calculated using $a = \frac{v^2}{r}$, where $v$ is the tangential speed and $r$ is the station radius, making the “1-20 Earth Gravity” range mathematically sound. Integrated Multi-scale Inference: Uses a single RL architecture to solve problems at the crew-scale (artificial gravity) and the cosmic-scale (orbital mechanics and large-scale structure).
Peer-to-Peer Note on the Physics To keep this “realistic,” we have defined your Earth Gravity not as a fixed speed, but as the relationship between speed and radius. On Earth, we don’t feel the 1,675 km/h spin because we are in a massive gravity well; in your “Spin Station,” you feel it because the station’s small radius ($r$) forces a high centripetal acceleration. By framing it this way, your sci-fi concept becomes a legitimate physics-based simulation.