Cosmic Ray Modulation: Force-Field Simulation and Machine-Learning Forbush-Decrease Forecasting

Galactic cosmic rays (GCRs) entering the heliosphere are modulated by the time-varying solar wind and heliospheric magnetic field, producing the well-documented 11-year (Schwabe) and 22-year (Hale) cyclic variation in cosmic ray intensity observed at Earth, together with transient, CME-driven depressions known as Forbush decreases (FDs). This paper (i) reviews the physical basis of cosmic ray modulation, centered on the force-field approximation of Gleeson and Axford and its use in reconstructing the solar modulation potential; (ii) reviews the growing literature applying machine learning and deep learning to cosmic ray intensity and Forbush-decrease forecasting; and (iii) reports two complementary, fully reproducible simulation-based experiments. First, a multi-solar-cycle (62-year) force-field simulator reproduces the characteristic anti-correlation between sunspot number and cosmic ray intensity (correlation coefficient r = -0.904) and the drift-related hysteresis loop between cosmic ray intensity and solar activity for opposite heliospheric magnetic polarity states. Second, an hourly-resolution, 3-year Forbush-decrease simulator, driven by the same class of solar wind turbulence proxies used in the companion geomagnetic-storm study, was used to train and evaluate Random Forest, Gradient Boosting, Multilayer Perceptron, and linear baseline models at 6-, 24-, and 72-hour forecast horizons under a chronologically leakage-aware protocol. The best model (MLP) achieved RMSE = 0.41% and R² = 0.965 at 6 hours, degrading to RMSE = 1.41% and R² = 0.592 at 24 hours, and to marginal skill by 72 hours, closely mirroring the multi-horizon skill-decay pattern reported for geomagnetic Dst forecasting and for real Forbush-decrease nowcasting studies in the literature. Feature-importance and ablation analysis confirm that recent cosmic-ray history and the local interplanetary magnetic field magnitude dominate short-horizon predictability. Because live access to real neutron monitor and OMNI archives was not available in this environment, both experiments are explicitly disclosed as physically-grounded simulations rather than analyses of observational data, and the paper concludes with a discussion of the sim-to-real gap, operational relevance (radiation dose forecasting, single-event upsets, atmospheric ionization), and directions for future work. Keywords: machine learning; neural networks; Random Forest; space weather; Heliosphere; cosmic ray modulation; force-field approximation; Forbush decrease; solar modulation potential