Causal learning across four space-environment regimes: autonomous discovery of regime-dependent physics and a measurement-gap roadmap for mission autonomy

Preprint. Distributed for review; comments welcome.

Abstract

Operational space-weather forecasting today relies on regime-specific models — NRLMSISE-00 for low-Earth-orbit (LEO) drag, AE9/AP9 for trapped particles at geosynchronous orbit (GEO), NAIRAS for Mars surface radiation, naive ballistic propagation from Lagrange-point 1 (L1) to the lunar distance for cislunar — that share little structure across regimes and produce point predictions without calibrated uncertainty. We present a per-edge causal-learning substrate, the Nervous Machine (NM) framework, and demonstrate its operational portability across four space-environment regimes. The same framework primitives, byte-identical across regimes, operate on each regime's archive data with no source-code changes: LEO thermospheric density against MSIS on 7.5 years of GRACE-FO accelerometer data; GEO energetic-particle and magnetic-field observables against the SWPC Relativistic Electron Forecast Model (REFM) on 7 days of GOES-19; Mars surface dose-rate on 1 Earth year of real MSL/RAD ground truth from the NASA Planetary Data System; and cislunar interplanetary magnetic field (IMF) at lunar distance against naive ballistic L1 propagation on the May 2024 G5 storm window from ARTEMIS-P1/P2. The substrate produces calibrated per-edge certainty (Z) and signed coupling (W) for every driver-observable-voxel triple, evolves them online from prequential residuals, and surfaces voxel-dependent physics regime structure without any voxel-specific prior. Anomaly-flag precision exceeds operational baselines or operational-physics naive comparators in three of four regimes; the cislunar comparator exhibits a voxel-dependent winner pattern that is itself the architecture's load-bearing falsifiable signature. More consequentially: the substrate independently discovered two regime-dependent physical phenomena from telemetry alone — magnetotail decoupling of the lunar-distance magnetic field from the upstream L1 driver, and the voxel-dependent Martian Forbush decrease in which storm-driven heliospheric disturbances suppress surface dose against operator intuition — without prior coding of either mechanism. This is the defining capability of a Physical AI: not curve fitting against known physics, but autonomous discovery of unmodeled physical interactions. The substrate as benchmarked here is the outer (environmental) learning loop of a two-loop assured-autonomy architecture; its converged per-edge priors seed an inner (mechanical) learning loop that runs on-payload to learn each spacecraft's own driver-observable structure. Together the two loops bound the operator's worst-case unknown unknowns from above (environment) and below (vehicle). We translate the residual Z-convergence gaps across the four regimes into a concrete missing-measurement prescription: which instruments, on which mission types, would close which gaps, prioritized by mission-autonomy criticality.

Keywords: space weather, Physical AI, causal learning, mission autonomy, regime portability, anomaly detection, calibrated uncertainty, on-payload inference.

Companion datasets cited in this paper

This paper analyzes four open-data regime benchmarks published as separate Zenodo records:

• LEO (GRACE-FO vs NRLMSISE-00, 7.5 yr): https://doi.org/10.5281/zenodo.20331839
• GEO (GOES-19 vs SWPC REFM, 7 d): https://doi.org/10.5281/zenodo.20563166
• Mars (MSL/RAD real ground truth, 1 yr): https://doi.org/10.5281/zenodo.20563381
• Cislunar (ARTEMIS-P1/P2, May 2024 G5 storm): https://doi.org/10.5281/zenodo.20563338

Source code: https://github.com/Nervous-Machine/cislunar-mars

Full draft available in the downloads (html)

Contact

heidi@everychart.io