Research/ ORBITAL ROBOTICS
Zar-Manah HELIOS 2.0: Dual-Horizon Architecture for Real-Time Space Weather & CME Arrival Forecasting
Zar Manah
Learning-based Methodstext match
Changes in v2.5.15: Documentation and metadata cleanup; no scientific changes. Geomagnetic superstorms driven by solar coronal mass ejections (CMEs) and high-speed solar wind streams pose catastrophic risks to continental power transmission grids, low-Earth-orbit (LEO) satellite constellations, and trans-polar aviation. While numerical magnetohydrodynamic (MHD) simulations and deep learning models have emerged as standard forecasting tools, operational space weather infrastructure has historically suffered from three critical bottlenecks: 1. **The Southward $B_z$ Paradox:** While coronagraphs measure launch kinematics to project arrival times ($24\text{--}72\text{ hours}$ in advance), determining whether the internal flux rope magnetic field points southward ($B_z \le -10\text{ nT}$, driving violent magnetic reconnection) or northward ($B_z > 0$, deflecting harmlessly) has required waiting for plasma to directly impact monitor spacecraft at Sun-Earth Lagrange Point 1 (L1), offering a mere $30\text{--}45\text{ minutes}$ of actionable warning and inducing false-alarm rates exceeding $70\%$. 2. **Computational Latency and the Riley Barrier:** Operational 3D numerical MHD models (e.g., NASA WSA-ENLIL) require hours of high-performance computing (HPC) cluster runtime and remain bounded by an arrival uncertainty of $\pm 10\text{--}12\text{ hours}$ due to single-viewpoint coronagraph projection ambiguities. 3. **Absence of Calibrated Finite-Sample Uncertainty:** Classical deterministic point predictions provide no rigorous error bounds for grid operators preparing for high-impact low-frequency (HILF) geomagnetic disasters. Here, we present **Zar-Manah HELIOS 2.0**, an integrated, four-tier space weather intelligence architecture executing locally on Apple Silicon M4 hardware with an end-to-end inference latency of **$16.55\text{ milliseconds}$** (median $16.54\text{ ms}$, $p99$ $17.34\text{ ms}$; sub-18 ms latency). HELIOS 2.0 unifies four synchronized operational gears across multiple spatial and temporal horizons: - **Gear 1 (Coronal Chirality & $B_z$ Threat):** Powered by `HELIOS-CHIRALITY-APEX`, evaluating southward magnetic threat ($B_z \le -10\text{ nT}$) and Inductive Conformal Quantile bounds $[Q_{10}, Q_{90}]$ directly at launch ($t = 0$). Evaluated out-of-distribution across unseen Solar Cycle 25 (91 CMEs) under forced uncalibrated binary classification (threshold 0.525), physical features achieve **$58.2\%$ accuracy**, **$0.6497$ ROC-AUC**, **$70.8\%$ specificity**, and **$44.2\%$ sensitivity** (TSS $+0.1502$). Crucially, 5-Fold Cross-Validation confirms a training baseline of **$0.4745$ ROC-AUC** (near chance), demonstrating that internal microscale flux rope magnetic reconnection 150 million km away cannot be reliably forecasted from early solar observations alone. Rather than deploying a pseudo-random coin toss under operational conditions, the deployed inference engine enforces a strict **selective classification / abstention policy ($0.40 < P_{\text{south}} < 0.60$)**. Under this policy, the system abstains on **$97.8\%$ of events (89 of 91)**, explicitly flagging them as `INDETERMINATE POLARITY [Abstain: Awaiting L1 In-Situ]` to prevent dangerous false security in power grid control rooms, requiring L1 in-situ confirmation (Gear 4) before declaring magnetic safety. - **Gear 2 (Severe Storm Veto):** Powered by `HELIOS-STORM-APEX`, synthesizing 3D de-projected CME kinetic energy ($E_{3\text{D}} = \frac{1}{2} M v_{3\text{D}}^2$) and soft X-ray flare fluence across 14 verified physical dimensions. Evaluated strictly out-of-distribution across unseen Solar Cycle 25 (2020–2026, 91 CMEs with verified Kyoto WDC causal arrival labels), `HELIOS-STORM-APEX` achieves **$78.0\%$ accuracy**, **$0.8737$ ROC-AUC**, and **$78.9\%$ specificity** (avoiding 60 out of 76 historical false alarms, with **$73.3\%$ sensitivity** capturing 11 of 15 severe storms, TSS $+0.5228$). With calibrated probabilities, it achieves a Brier score of **$0.1154$**, strictly outperforming climatological baseline ($0.1377$). - **Gear 3 (Interplanetary Shock Kinematics):** Powered by `MoKTransitNN` and a dynamic Gradient Boosted decision tree regime trained on 331 verified empirical CME shocks without synthetic leakage. Evaluated out-of-distribution across unseen Solar Cycle 25 (91 clean CMEs), HELIOS achieves an arrival MAE of **$10.63\text{ hours}$** and a median absolute error of **$7.45\text{ hours}$** (GBDT alone) and $11.69\text{ hours}$ MAE ($8.34\text{ hours}$ median, MoK hybrid ensemble), approaching the operational literature standard reported for NASA's WSA-ENLIL 3D MHD model on Solar Cycle 24 ($\pm 10.40\text{ hours}$ at Earth; Wold et al., 2018; Riley et al., 2018) in sub-second inference on edge hardware ($<0.05\text{ s}$), alongside finite-sample inductive conformal prediction intervals ($92.3\%$ empirical coverage on nominal 90% interval, mean width $59.7\text{ h}$, satisfying the nominal $90\%$ finite-sample guarantee without calibration leakage). *Note on benchmark context: Wold et al. (2018) evaluated a historical sample from 2010–2016, whereas HELIOS is evaluated on Cycle 25 (2020–2026).* - **Gear 4 (Tactical L1 In-Situ Attention):** Synchronizes two complementary models on real-time solar wind streams: (1) `HELIOS-Point` (GBDT + CQR), providing multi-horizon $Dst$ forecasts across $+1\text{h}$ to $+6\text{h}$ on all 57,998 samples of Solar Cycle 25 ($R^2 = 0.968$, RMSE $= 3.66\text{ nT}$, MAE $= 2.36\text{ nT}$ at $+1\text{h}$; $R^2 = 0.714$, RMSE $= 10.91\text{ nT}$ at $+6\text{h}$, with $89.26\%$ finite-sample calibrated CQR coverage on nominal 90% interval, $86.43\%$ raw quantile coverage); and (2) `HeliosBiLSTM`, a 14-feature bidirectional recurrent neural network with temporal attention providing standalone $+1\text{h}$ neural nowcasts directly on Apple Silicon M4 GPU ($R^2 = 0.955$, RMSE $= 4.32\text{ nT}$, MAE $= 2.54\text{ nT}$ across 57,975 rolling 24h sequences). We integrate an operational client streaming live telemetry from NOAA SWPC and NASA SDO/SOHO. In operational tests on Apple Silicon M4, HELIOS 2.0 ingested live satellite feeds, executing all four tiers in **$16.55\text{ ms}$** (median $16.54\text{ ms}$, $p99$ $17.34\text{ ms}$), correctly identifying active high-speed streams as northward and non-hazardous, defusing false alerts. All checkpoints, scalers, and reproducible code are maintained open-source under the MIT license for operational research and open scientific benchmarking. **Keywords:** Space Weather, Geomagnetic Storms, Dst Index, Southward Bz Prediction, Coronal Mass Ejections, Conformal Prediction, Mixture-of-Kinematics, Apple Silicon M4. --- Changes in v2.5.14 (supersedes v2.5.11): Published PDFs, metrics and code are now byte-consistent with the GitHub tag v2.5.14; the v2.5.11 deposit mixed 81.3% and 92.3% transit-coverage figures in different files. Model checksum manifest regenerated and verified: every sha256 in helios_models_checksums.json now validates its file. Ablation table clarified: the PHYS14 row is an ExtraTrees reference, not the deployed LightGBM APEX estimator; both score near chance under cross-validation. Tactical coverage/width are now reproducible from the published models and feature parquet via scripts/20_verify_tactical_metrics.py, which regenerates test-prediction parquets for all four horizons. Latency benchmark artifact refreshed with measurement protocol and run-to-run variance note. predict.py --live now reports data_source and fallback_reason in JSON, so a synthetic fallback can never be mistaken for live data. --dashboard opens the cockpit correctly on macOS; --version reports 2.5.14. Changes in v2.5.14 (supersedes v2.5.12): Correspondence now points to the repository issue tracker. No other scientific content, metric or result changed with respect to v2.5.12.