Compartmentalized Diffusion Models (CDMs)

• Compartmentalized Diffusion Models are frameworks that describe transport in subdivided media by modeling intra-compartment stochastic dynamics and inter-compartment exchanges across barriers.
• They reveal how microscopic properties such as barrier permeability and compartment size connect with macroscopic behaviors like anomalous diffusion and non-Gaussian displacement distributions.
• CDMs are applied in biophysical membrane analysis, porous media studies, diffusion MRI, and modular generative modeling, offering robust insights into inverse problems and transport phenomena.

Compartmentalized Diffusion Models (CDMs) encompass a family of theoretical and computational frameworks for modeling transport processes in media subdivided into distinct regions or “compartments,” each separated by barriers that restrict or mediate the movement of diffusing agents, particles, or information. In both physical systems (e.g., biological membranes, disordered media) and machine learning applications (e.g., modular generative models), compartmentalization introduces rich dynamical behaviors, nontrivial transport statistics, and unique algorithmic properties. CDMs have been developed across statistical physics, stochastic simulations, diffusion MRI, and computational generative modeling, each domain leveraging the notion of local dynamics and inter-compartmental exchange governed by interface properties.

1. Stochastic Models of Physical Compartmentalized Diffusion

Stochastic compartmentalized diffusion models describe a single or ensemble of particles diffusing in a spatially heterogeneous environment, partitioned by semi-permeable or partially transmitting barriers. A canonical model consists of a one-dimensional domain segmented into intervals (“compartments”) with lengths ${L_k}$ drawn i.i.d. from an exponential or heavy-tailed law, separated by barriers of finite permeability $\kappa$ or transmittance $T$ (Ślęzak et al., 2019, Muñoz-Gil et al., 2018).

Within each compartment, particles perform free Brownian motion with diffusivity $D$. At each barrier, particles are either transmitted (escape to the neighboring compartment) or reflected, with probabilities defined by the permeability/transmittance of the barrier:

• Escape dynamics: The first-passage time to cross a barrier is typically exponentially distributed with rate proportional to $2\kappa D/L$ where $L$ is the current compartment length (Ślęzak et al., 2019).
• Jump rules (CTRW interpretation): Upon escape, the realized jump is to a uniformly selected position in the adjacent compartment, with jump length $J_n = \pm L_n / 2$, and inter-event waiting time $T_n = E_n L_n / 2$ for $E_n \sim \mathrm{Exp}(1)$.

Microscale and mesoscale descriptions are formally equivalent for long-time statistics (Muñoz-Gil et al., 2018), with mesoscale models interpreting the process as a sequence of “flights” (crossing events) interleaved by stochastic sojourns.

2. Statistical Properties and Anomalous Transport

The principal statistical observables in CDMs include the propagator $P(x, t)$ (probability density at position $\kappa$0 at time $\kappa$1), the mean-squared displacement (MSD) $\kappa$2, time-dependent diffusion coefficient $\kappa$3, and higher moments such as kurtosis $\kappa$4.

Key results:

• Non-Gaussian PDF: In the transient regime ($\kappa$5, with $\kappa$6 and $\kappa$7), the propagator $\kappa$8 exhibits sharp exponential (Laplace) tails, $\kappa$9, with strong deviation from Gaussianity (Ślęzak et al., 2019). The excess kurtosis $T$0 decays from 3 (Laplace) toward 0 (Gaussian) as time increases.
• Brownian yet non-Gaussian: Linear MSD scaling ($T$1) coexists with non-Gaussian propagators at intermediate times; only in the long-time ($T$2) and very short-time ($T$3) limits does the process revert to Gaussian statistics at distinct diffusivities.
• Anomalous diffusion:
  • Pure geometric disorder (random $T$4 with fixed $T$5) yields normal diffusion at long times.
  • Disorder in barrier transmittances ($T$6 heavy-tailed) is required for subdiffusion, with scaling $T$7 for $T$8.
  • Coincidence or competition of geometric and transmittance disorder yields regimes where spatial heterogeneity can mitigate or suppress subdiffusion (Muñoz-Gil et al., 2018).

This taxonomy provides a framework connecting microscopic features (barrier distribution, spatial disorder) with macroscopic transport exponents and spectrum of possible diffusive behaviors.

3. Applications in Physical, Biological, and Medical Domains

CDMs are foundational for interpreting transport phenomena in a wealth of systems:

• Biophysics/Cellular Membranes: Hop diffusion of single particles in the plasma membrane, where the characteristic sharp central peaks and exponential tails of $T$9 have been directly observed; measured values for the effective diffusivity ratio $D$0 constrain microscopic parameters such as mesh size and barrier permeability (Ślęzak et al., 2019).
• Porous and Composite Media: CDMs explain non-Gaussian displacement distributions in soft composites and porous rocks; two-scale Fickian behavior links local geometric heterogeneity and macroscopic spreading.
• Diffusion MRI: Multi-compartmental models such as SANDI, VERDICT, and CEXI map signal attenuation to microstructural parameters (e.g., cell size, intracellular volume fraction, permeability). For permeable tissue, only models including explicit trans-compartment exchange (e.g., CEXI) provide unbiased microstructure estimation or accurate permeability/exchange rates (Gardier et al., 2022).

A summary table of three prominent diffusion MRI models:

Model	Exchange Included	Key Fitted Parameters
SANDI	No	$D$1, $D$2, $D$3, $D$4, $D$5
VERDICT	No	$D$6, $D$7, $D$8, $D$9
CEXI	Yes	$2\kappa D/L$0, $2\kappa D/L$1, $2\kappa D/L$2, $2\kappa D/L$3, $2\kappa D/L$4

In permeable systems, the inclusion of inter-compartment exchange is essential for robust parameter recovery.

4. Pseudo-Compartment and Hybrid Modeling Approaches

Hybrid simulation of reaction-diffusion systems at multiple scales employs the pseudo-compartment method (Yates et al., 2015). This approach rigorously couples a PDE-based continuum description (for regions of high copy number) with a compartment-based stochastic model (for regions of low copy number) via a special “pseudo-compartment” spanning the interface:

• In the compartment region, particles are tracked as discrete stochastic entities in a lattice.
• In the PDE region, concentrations are evolved via deterministic partial differential equations.
• Exchange between regions is mediated by stochastic jumps and mass transfer with exact (Algorithm 1) or continuum-limit (Algorithm 2) coupling, preserving particle number and fluctuations across the interface.

This scheme is robust to low/high copy number, avoids complications of flux-balancing or overlap regions, and applies to generalized geometries and higher dimensions (Yates et al., 2015).

5. Compartmentalized Diffusion Models in Generative Machine Learning

An independent codification of “Compartmentalized Diffusion Models” has emerged in generative modeling, denoting a modular framework where diffusion models are trained separately on mutually disjoint data partitions (“compartments”) and composed at inference time (Golatkar et al., 2023):

• Data $2\kappa D/L$5 is split into $2\kappa D/L$6 disjoint compartments $2\kappa D/L$7.
• Independent diffusion submodels $2\kappa D/L$8 are trained on each $2\kappa D/L$9, yielding $L$0 generative modules with no crosstalk.
• At sampling time, the global score is formed as $L$1; the weights $L$2 are posterior probabilities reflecting the likelihood that a perturbed sample arises from each compartment.

Crucial properties include:

• Perfect selective forgetting: Any compartment can be unlearned by excising its submodel, strictly removing all traces of that data.
• Continual learning: New data partitions can be incorporated by training new submodels, avoiding catastrophic forgetting.
• Attribution and modular access: The influence of training partitions on generated samples is precisely quantifiable.

Empirical results demonstrate near-paragon generation quality (class-conditional FID within 10% of non-compartmentalized models for up to 8 splits), rapid unlearning (8x faster for 8-way splits), improved text-image alignment (TIFA), and enhanced robustness to memorization attacks (Golatkar et al., 2023).

6. Theoretical Synthesis and Outlook

Compartmentalized Diffusion Models unify disparate domains by formalizing the interplay between intra-compartment stochasticity and inter-compartmental exchange, whether that exchange is physical (semi-permeable barriers; $L$3, $L$4), algorithmic (modular model composition), or reaction-limited (diffusion barriers in tissue microstructure). Across contexts, key themes emerge:

• The statistics of dwell times and inter-barrier jumps determine macroscopic transport properties.
• Disorder (in compartment size or barrier permeability) can either induce or mitigate anomalous diffusion and must be quantitatively distinguished (Muñoz-Gil et al., 2018).
• Inverse problems in imaging require explicit modeling of exchange across compartments to attain accurate microstructure estimation (Gardier et al., 2022).
• In machine learning, strict parameter partitioning and closed-form compositionality enable previously unattainable guarantees for data unlearning and modular continual learning (Golatkar et al., 2023).

Current research explores extension to multi-scale and multi-regime hybrids, generalization to multi-dimensional geometries, and the integration of CDMs with advanced generative and inverse modeling pipelines across scientific and engineering disciplines.