BELM-MDCM: Astrophysics & Causal Inference Model

• BELM-MDCM is a comprehensive framework that extends the BELM code to simulate stochastic particle acceleration in high-energy astrophysical coronae using coupled kinetic and MHD modeling.
• It implements a diffusion-based causal generative model that enforces analytically invertible mappings and the Causal Information Conservation principle for deterministic counterfactuals.
• Empirical results validate its performance by replicating spectral observations in accretion disk coronae and enhancing causal inference accuracy on benchmark datasets.

BELM-MDCM denotes a specialized framework or extension built upon the BELM code family, implemented in diverse domains including particle acceleration in black hole coronae and causally faithful generative modeling via diffusion processes. Across its principal technical instantiations, BELM-MDCM emphasizes first-principles kinetic or probabilistic modeling, energy or information conservation, and numerically robust treatment of coupled particle–wave or data-generating mechanisms.

1. Formalism and Domain Definitions

In high-energy astrophysics, BELM-MDCM generalizes the BELM (Boltzmann–Electron–Lepton–Magnetized) code to simulate stochastic particle acceleration in accretion disk coronae, integrating a full magnetohydrodynamic (MHD) wave cascade with self-consistent coupling to thermal protons and leptons. The model resolves the interplay between turbulent slab waves, particle populations (protons and $e^\pm$), and radiative processes through a set of coupled kinetic equations—chiefly, a wave energy density PDE with injection, cascade, and resonance-damping terms, and a Fokker–Planck equation governing the evolution of particle momentum distributions subject to both stochastic acceleration and losses.

In structural causal modeling, BELM-MDCM refers to a diffusion-model-based generative architecture enforcing analytically invertible mappings to eliminate structural reconstruction error (SRE), thereby guaranteeing deterministic and lossless abduction under the Causal Information Conservation (CIC) principle. Each system variable is assigned either a full BELM-based causal diffusion generator or a simpler mechanism, coupled through the DAG of the structural causal model, trained with a hybrid loss enforcing causal and generative fidelity (Wu et al., 7 Nov 2025).

2. Mathematical Framework

Astrophysical BELM-MDCM

The hot corona above the accretion disk is modeled as a sphere of radius $R$ containing:

• Thermal protons ($T_p$)
• Soft photons (disk blackbody at $T_{\rm bb}$)
• Energetic leptons ($e^\pm$)
• Magnetized slab MHD fluctuations (two polarizations: Left "L", Right "R")

The model encompasses:

• Wave Dynamics: Energy density $W_\sigma(k,t)$ of each mode $\sigma$ is evolved under

$$\frac{\partial W_\sigma}{\partial t} = \frac{\partial}{\partial k}\left[D_{kk}(k)\frac{\partial W_\sigma}{\partial k}\right] - \alpha_\sigma(k)W_\sigma + S_\sigma(k)$$

with non-linear diffusion $D_{kk}$, injection $S_\sigma$, and damping $R$0 computed from the resonance kernel $R$1.

• Gyroresonant Particle–Wave Coupling: Energy/momentum exchange occurs when the resonance condition

$R$2

is satisfied ($R$3), enabling stochastic acceleration.

• Particle Kinetics: Each species’ momentum distribution $R$4 evolves via:

$R$5

where $R$6 is momentum diffusion from all waves, and $R$7 includes both radiative and Coulomb losses.

Diffusion-based BELM-MDCM for Counterfactuals

Each SCM node with parents $R$8 is modeled as:

• Abduction: $R$9 (ENCODE)
• Generation: $T_p$0 (DECODE)

The BELM mechanism ensures $T_p$1, formally yielding $T_p$2 for all $T_p$3. The hybrid objective balances score-matching ($T_p$4) and causal task ($T_p$5) losses, enforcing both model fit and causal identifiability.

3. Numerical Implementation and Algorithmics

BELM-MDCM in Accretion Coronae

• Grid Discretization: $T_p$6-space (256 points, $T_p$7 to $T_p$8); $T_p$9-space (128 points, $T_{\rm bb}$0 to $T_{\rm bb}$1).
• Time-Stepping: Semi-implicit for waves (implicit in diffusion, explicit in damping); Crank–Nicolson for Fokker–Planck (diffusion), explicit for advection/loss.
• Boundary Conditions: Injection at $T_{\rm bb}$2, free outflow at $T_{\rm bb}$3; reflecting at $T_{\rm bb}$4, free escape at $T_{\rm bb}$5.
• Loop: (i) Compute absorption/acceleration rates, (ii) advance wave and particle fields, (iii) update photon field, (iv) check energy conservation.

Convergence to steady state typically requires $T_{\rm bb}$6 timesteps with $T_{\rm bb}$7 (Marcowith et al., 2013).

BELM-MDCM for Causal Counterfactuals

• Targeted Modeling: Only SCM nodes central to causal queries (e.g., $T_{\rm bb}$8 and $T_{\rm bb}$9) receive costly diffusion solvers; background nodes may use reduced models.
• Abduction/Prediction Workflow: Given observation, encode to latent $e^\pm$0 via analytically invertible BELM encoder; apply intervention; decode to counterfactual $e^\pm$1 via BELM decoder.
• Loss Minimization: Hybrid loss for each node: $e^\pm$2.
• Pseudocode: See Algorithm 1 in (Wu et al., 7 Nov 2025).

4. Regulatory Mechanisms and Physical Effects

Proton-Temperature Regulation ("Proton Switch") in Coronae

Protons can efficiently absorb turbulent power, quenching electron acceleration above a threshold:

$e^\pm$3

• If $e^\pm$4: almost all turbulent energy cascades to protons; electrons remain thermal.
• If $e^\pm$5: R-modes accelerate $e^\pm$6; hybrid distributions emerge (Marcowith et al., 2013).

This self-regulates spectral states and may underlie transitions in X-ray binaries and AGN.

CIC Principle in Counterfactuals

CIC ("Causal Information Conservation") dictates that abduction and generation must be exact inverses: $e^\pm$7 for all $e^\pm$8. This is necessary for logically precise counterfactuals and is incorporated by construction in BELM-MDCM's analytically invertible diffusion architecture (Wu et al., 7 Nov 2025).

5. Empirical Results and Applications

High-Energy Astrophysics

• Preliminary application to the high-soft state (HSS) of Cygnus X-1 yields best-fit parameters: $e^\pm$9 cm, $W_\sigma(k,t)$0, $W_\sigma(k,t)$1, $W_\sigma(k,t)$2, $W_\sigma(k,t)$3, $W_\sigma(k,t)$4 $W_\sigma(k,t)$5, $W_\sigma(k,t)$6 keV. Model output includes a dominant thermal Compton peak ($W_\sigma(k,t)$7–$W_\sigma(k,t)$8 keV) and a soft nonthermal tail ($W_\sigma(k,t)$9 up to $\sigma$0 MeV). The model achieves reduced $\sigma$1, demonstrating quantitative agreement with observed spectra (Marcowith et al., 2013).

Causal Inference

• On highly nonlinear confounding, BELM-MDCM achieves ATE error $\sigma$2 (true ATE: $\sigma$3), outperforming classical propensity models.
• On Lalonde benchmark: error $\sigma$4 vs. Causal Forest's $\sigma$5.
• In stress tests on non-invertible SCMs, BELM (zero SRE) yields PEHE $\sigma$6; for classical DDIM PEHE increases to $\sigma$7.
• Ablation indicates performance degrades if analytical invertibility (BELM), hybrid loss, or targeted modeling are omitted (Wu et al., 7 Nov 2025).

6. Limitations and Prospects

• Astrophysics: Present implementation is restricted to 1D slab turbulence with only slab modes; future extensions include transit-time damping and multidimensional wave spectra. Low-hard state modeling and AGN coronae are active areas for extension (Marcowith et al., 2013).
• Causal Modeling: BELM-MDCM assumes invertible mechanisms and correct input DAGs; adaptation to many-to-one SCMs and graph-uncertainty remains to be developed. Scaling to high-dimensional data remains computationally demanding (Wu et al., 7 Nov 2025).

7. Synthesis and Impact

BELM-MDCM formalizes a paradigm of self-consistent, energetically and/or informationally closed dynamic models applicable in both microphysical (astrophysical plasma) and causal-inference (generative modeling) settings. Its distinctive features are physically transparent handling of coupled nonlinear processes, explicit regulatory mechanisms (e.g., proton switch, CIC), and empirically validated, robust algorithmic implementations. In both domains, BELM-MDCM establishes a new standard for predictive precision and interpretability in systems controlled by turbulent or latent-variable-driven stochasticity (Marcowith et al., 2013, Wu et al., 7 Nov 2025).