Immiscible Diffusion: Models & Applications

• Immiscible diffusion is the process where molecular exchange occurs primarily at interfaces in systems with thermodynamically immiscible components.
• Mathematical models like the Navier–Stokes/Cahn–Hilliard system capture the interfacial transport and kinetic behaviors that arise from suppressed bulk mixing.
• Applications range from microfluidics, polymer physics, and alloy thin films to machine learning, demonstrating the wide impact of controlled interfacial diffusion.

Immiscible diffusion describes diffusion processes in systems where distinct components are thermodynamically immiscible, such that macroscopic mixing is suppressed but molecular or mesoscopic exchange persists at interfaces or under specific conditions. The phenomenon arises in diverse contexts, including multicomponent fluid dynamics, polymer physics, crystalline solids, microfluidics, alloy thin films, and machine learning models. Fundamental to immiscible diffusion is the interplay between suppressed bulk intermixing and localized, interfacial or geometry-driven transport, producing distinct kinetic, morphological, and dynamic behaviors.

1. Mathematical Formulations and Physical Models

In continuum fluid mechanics, immiscible diffusion is rigorously described by diffuse-interface models coupling transport and hydrodynamics. Prototypical formulations include the incompressible Navier–Stokes/Cahn–Hilliard (NS/CH) system with nonlocal interactions and nonconstant mobility (Frigeri et al., 2013). The primary variables are the velocity field $u(x,t)$, pressure $\pi(x,t)$, order parameter %%%%2%%%% (relative concentration, $\phi\in(-1,1)$), and chemical potential $\mu(x,t)$. The governing equations read: $$\begin{aligned} & u_t - \nu \Delta u + (u \cdot \nabla)u + \nabla \pi = \mu \nabla \phi + h, \qquad \nabla \cdot u = 0,\ & \phi_t + u \cdot \nabla \phi = \nabla \cdot [m(\phi) \nabla \mu], \ & \mu = a(x) \phi - J * \phi + F'(\phi), \end{aligned}$$ where $m(\phi)$ is the (possibly degenerate) mobility, $J$ is the nonlocal interaction kernel, $F$ is the double-well or singular potential, and $*$ denotes convolution.

Degenerate mobility ($m(\pm1) = 0$) pins pure phases to inhibit bulk diffusion, while nonconstant $m(\phi)$ localizes diffusion to interfacial regions. Nonlocal coupling via $J$ introduces long-range correlations, supplementing classical gradient flows. In the compressible context, quasi-incompressible and varying-density extensions (Lowengrub–Truskinovsky model) include $$\rho(c) \partial_t c + \rho(c) v \cdot \nabla c = \nabla \cdot (m(c) \nabla \mu)$$ with pressure-coupled chemical potential, mass balance, and stress tensor modifications (Abels, 2011).

These frameworks guarantee global existence, energy dissipation, and attractors under well-posedness theory, and allow for physically accurate interface migration, morphology selection, and kinetic arrest in immiscible two-phase flows (Frigeri et al., 2013, Chen et al., 2021).

2. Interfacial Diffusion in Immiscible Systems

In immiscible systems, diffusion is confined to interfaces or driven by defects, geometry, or external fields rather than homogeneous intermixing. At the interface, the chemical potential gradient $\nabla \mu$ can be large while mobility $m(\phi)$ is nonzero only near $\phi \approx 0$, producing sharply localized fluxes $J_\text{diff} = -m(\phi)\nabla \mu$.

In polymeric materials, molecular dynamics show that increasing immiscibility drastically arrests interdiffusion, limiting the equilibrium interfacial width to $w_\text{eq} \sim a/\sqrt{\chi}$ ($\chi$: Flory–Huggins parameter), and precluding the formation of topological entanglements essential for mechanical strength (Ge et al., 2013). When $w_\text{eq} < w_c \approx 6a$, there is essentially zero interfacial entanglement density, and mechanical failure occurs by chain pullout at low stress.

In alloy thin films, local film–substrate reactions induce immiscible phase separation and coupled diffusion pathways. For instance, in Ag–Cu/Si systems, Cu diffuses laterally through an Ag-rich halo to a local reaction front, depleting Cu near substrate features, with the interfacial diffusivity orders of magnitude higher along grain boundaries (GB-mediated diffusion) (Peddiraju et al., 10 Jan 2026). The resulting microstructure deviates from classical diffusion growth (Stefan-type laws), producing kinetic exponents $0.29$–$0.46$ distinct from the standard $1/2$.

3. Immiscible Diffusion in Microfluidics and Confinement

Microdroplet dissolution in sparsely miscible environments showcases the influence of confinement, wall permeability, and collective effects. For water droplets in silicone oil confined by PDMS walls, dissolution is primarily controlled by diffusion through the continuous phase, modeled by the Epstein–Plesset law: $$R^2(t) = R_0^2 - 2 \frac{D_\mathrm{o}}{\rho} (C_s - C_\infty) t,$$ where $D_\mathrm{o}$ is water diffusivity in oil, and $C_s$ is saturation concentration (Zhang et al., 2020). However, permeability of PDMS introduces an additional vapor sink, accelerating dissolution beyond classical theory. For sufficiently large droplets or droplet clusters, finite oil solubility and neighbor screening reduce dissolution rates, demanding full numerical solutions.

Confinement ratio ($c = R_0/R_B$) and neighbor spacing ($L/D$) sharply affect lifetimes, highlighting that even minimal miscibility enables significant interfacial transport in otherwise immiscible systems if geometry or boundary conditions support it.

4. Immiscible Diffusion in Crystalline Solids and Nanostructures

Spinodal decomposition and self-organization in immiscible crystalline materials, such as PbTe/CdTe, result from the synergistic action of bulk and surface diffusion, anisotropic mobilities, and thermodynamic instability (Mińkowski et al., 2016). Kinetic Monte Carlo simulations demonstrate that enhanced bulk diffusion at elevated temperatures, along with a bias in vertical mobility, produces morphologies ranging from quantum dots (0D) to nanowires (1D) and complete phase-separated slabs (3D).

The kinetic modeling relies on

$$w_{i \rightarrow j} = \nu_0 \exp(-B/k_B T) \exp[-(\Delta E + k \Delta z/a)/(k_B T)]$$

and

$L_i(t) = \sqrt{D_i t}$

where $B$ is the diffusion barrier, $k$ the bias, and $D_i$ the effective diffusion coefficient. These results are sensitive to the balance between thermodynamic driving forces and anisotropic kinetics. Immiscibility enforces sharp interfaces and phase selectivity, but does not preclude nanoscale diffusion along specific pathways (e.g., surface, GB) under growth or annealing protocols.

5. Immiscible Diffusion in Machine Learning: Diffusion Models

In generative machine learning, “Immiscible Diffusion” refers to modified training strategies in denoising diffusion probabilistic models (DDPMs) that reduce random mixing between distinct data trajectories in noise space (Li et al., 2024, Li et al., 24 May 2025). Standard (miscible) training randomly pairs images and Gaussian noise, causing overlapping diffusion trajectories and ambiguous denoising targets. Immiscible diffusion introduces a batch-wise linear assignment (or alternative mechanisms such as nearest-neighbor noise assignment and image scaling) to preferentially couple images with proximate noise vectors.

This methodology increases the separation (in $L_2$ space) between images’ noisy representations, simplifies the denoising objective, and accelerates convergence (up to 4x across models and datasets). The benefit is quantified by the increase in the average centroid distance $D̄$ between clusters in noise space, with empirical $D̄$ rising from $\approx 0.92$ (vanilla) to $4.11$ (immiscible) in CIFAR-10 (Li et al., 24 May 2025). The process acts analogously to an “external force,” reducing entropy in the denoising mapping and preserving generative diversity by enforcing near-bijective correspondence between noise and data.

Implementation is inexpensive (one line of code), inference unaffected, and the approach is complementary to other architectural innovations. Extensions beyond computationally expensive linear assignment include KNN-based sampling and multiplicative image rescaling, all aiming to reduce the miscibility of data trajectories in latent space.

6. Numerical Methods for Immiscible Multiphase Flows

Simulating immiscible diffusion in high density-ratio, multi-material flows demands regularization schemes that maintain sharp interfaces, stability, and thermodynamic equilibrium. Standard localized artificial diffusivity (LAD) approaches, which target mass-fraction diffusion, become unstable at extreme density ratios and unsuited for immiscible interfaces. Enhanced methods employ:

• Artificial diffusion targeted on volume-fraction gradients or “ringing”
• Bulk-density diffusion for consistent pressure–temperature equilibrium
• Conservative diffuse-interface (“sharpening”) flux to enforce prescribed interface thickness (Brill et al., 16 Mar 2025)

Formally, the species-mass flux for component $i$ comprises: $$\mathbf{J}_i^* = -D^*_\rho(\mathbf{x}, t)\nabla(\rho Y_i)$$ and a sharpening flux

$$\mathbf{J}_{\text{sharp}, i} = -\rho_i \Gamma [\epsilon \nabla V_i - \sum_{j\neq i} V_i V_j \hat{n}_{ij}]$$

($\Gamma$ is sharpening speed, $\epsilon$ sets width). These ensure robustness, interface control, and conservation for arbitrary $N$-material flows, yielding stable evolution up to density ratios $10^6$ and close agreement with experimental shock–bubble phenomena.

7. Analytical, Computational, and Experimental Perspectives

Comprehensive analysis of immiscible diffusion spans:

• Rigorous global existence, attractor theory, and energy dissipation for diffuse-interface PDEs (Frigeri et al., 2013, Chen et al., 2021)
• Quantitative matching of diffusion kinetics and microstructure via inverse optimization in reaction–diffusion systems (Peddiraju et al., 10 Jan 2026)
• Molecular-scale simulation of interfacial arrest, stress–structure relations, and entanglement thresholds (Ge et al., 2013)
• High-order, multidimensional simulations of mixing suppression and interface regularization strategies (Brill et al., 16 Mar 2025)
• Empirical measurement and fit of droplet dissolution in confined, sparsely miscible systems (Zhang et al., 2020)
• Benchmarking and acceleration of learning curves in diffusion-model training via spatial decorrelation (Li et al., 2024, Li et al., 24 May 2025)

These complementary methodologies reveal that immiscible diffusion frequently controls the macroscopic kinetics, mechanical integrity, or algorithmic efficiency of multiphase systems—emphasizing the need for physically motivated, interface-aware models in both natural and artificial contexts.