← Back

Notes · phase separation with a passive species

Self-Propelled Cahn–Hilliard Droplets

A phase-separating field coupled to a passive, purely diffusing field by cross-diffusion. When the coupling is non-reciprocal, the droplets push the passive matter into a bow wave and chase it across the plane.

27 October 2026

1The Cahn–Hilliard equation

The Cahn–Hilliard equation models phase separation of a binary mixture with a conserved order parameter \(c\). It is a gradient flow of the Ginzburg–Landau free energy

\[ \begin{equation} F[c] = \int_\Omega \Big( f(c) + \tfrac{\varepsilon^2}{2}|\nabla c|^2 \Big)\, dV, \qquad f(c) = \tfrac14 (c^2-1)^2 \tag{1}\label{eq:F} \end{equation} \] \[ \begin{equation} \frac{\partial c}{\partial t} = \nabla\cdot\big(M\nabla\mu\big), \qquad \mu = \frac{\delta F}{\delta c} = f'(c) - \varepsilon^2\nabla^2 c \tag{2}\label{eq:CHgen} \end{equation} \]

With constant mobility this is the familiar form

\[ \begin{equation} \frac{\partial c}{\partial t} = M\,\nabla^2\big(c^3 - c - \varepsilon^2\nabla^2 c\big). \tag{3}\label{eq:CH} \end{equation} \]

2Coupling to a passive field

A second density \(\rho\) that only diffuses is trivially decoupled unless something links it to \(c\). Three common links:

  1. Free energy. Add an ideal-solution term and a cross term, \(F \to F + \int [k_BT\,\rho(\ln\rho-1) + \chi c\rho]\,dV\). Then \[ \begin{equation} \partial_t c = M\nabla^2\big(c^3 - c - \varepsilon^2\nabla^2 c + \chi\rho\big), \qquad \partial_t\rho = D\nabla^2\rho + \tfrac{D\chi}{k_BT}\nabla\cdot(\rho\nabla c). \tag{4}\label{eq:partition} \end{equation} \] The passive species partitions into one phase or onto interfaces, as client molecules do in condensates.
  2. Kinetics only. \(D = D(c)\) or \(M = M(\rho)\), as in crowded media.
  3. Reactions. Source and sink terms that convert or consume species, as in phase-field tumour models and chemically active droplets (Section 7).

3The cross-diffusion model

The system studied here couples a Cahn–Hilliard field \(A\) to a passive field \(P\) through four diffusion constants. It is linear apart from the cubic and \(\kappa\) terms:

\[ \begin{equation} \frac{\partial A}{\partial t} = \nabla^2\Big[ D_A\big(A^3 - A - \kappa\nabla^2 A\big) + D_{AP}\,P \Big] \tag{5}\label{eq:A} \end{equation} \] \[ \begin{equation} \frac{\partial P}{\partial t} = \nabla^2\Big[ D_P\,P + D_{PA}\,A \Big] \tag{6}\label{eq:P} \end{equation} \]

When a free energy exists

Writing each equation as \(\partial_t\phi_i = D_i\nabla^2\mu_i\), the cross terms come from a single \(\chi AP\) in a free energy only if

\[ \begin{equation} D_{AP}\,D_P = D_{PA}\,D_A. \tag{7}\label{eq:recip} \end{equation} \]

Then the dynamics is a gradient flow and droplets can only sit still and coarsen. When Eq. \eqref{eq:recip} fails the coupling is non-reciprocal and there is no Lyapunov functional.

Linear stability

About a uniform state \((A_0, P_0)\), with \(a = 3A_0^2 - 1\) and perturbations \(\propto e^{ikx+\lambda t}\), the growth rates are eigenvalues of

\[ \begin{equation} J(k) = -k^2\begin{pmatrix} D_A(a+\kappa k^2) & D_{AP} \\ D_{PA} & D_P \end{pmatrix}. \tag{8}\label{eq:J} \end{equation} \]
t = 0 ⟨A⟩ = 0 grid 128², Δt 0.1 A P
Figure 1. Live solution of Eqs. 5–6 with DA = κ = 1 on a periodic 128 × 128 box. Red dots: A. Blue dots: P. Dot density follows local concentration. Pseudo-spectral, semi-implicit: κ∇⁴ and DP∇² implicit; A³ and cross terms explicit with a stabilizing term. The readout ⟨A⟩ stays fixed because the k = 0 mode never changes.

4Why the droplets move

With \(D_{AP} = -\alpha\) and \(D_{PA} = +\alpha\):

A symmetric rim pulls equally in all directions. If it tips, the droplet moves toward the thicker side, pushes P ahead of itself like a bow wave, and leaves a depleted wake that refills only by diffusion, on a time of order \(R^2/D_P\). Small \(D_P\) preserves the asymmetry, so the droplet keeps chasing a pile it is building. Large \(D_P\) smooths the pile into a halo and motion stops.

Eqs. 5–6 are invariant under \(P\to-P,\ \alpha\to-\alpha\). The opposite sign convention gives droplets led by a low-P region; the motion is the same. Only \(D_{AP}D_{PA} \lt 0\) matters.

αDPSpeed (cells per unit time)Shape
0.100.050.011round
0.120.030.045ovals, some short worms
0.200.020.070elongated
0.250.020.10crescents and worms
Table 1. Droplet drift speed from offline sweeps (128² grid, Δt = 0.1, seeded droplets). Faster droplets stretch perpendicular to their motion.

The simulation advances about 900 steps per second in a browser regardless of the steps-per-frame setting, which only trades frame rate for steps. To go faster, raise α or lower \(D_P\), or change the code to use a larger Δt or a smaller grid.

5Why small droplets disappear

Total A and total P are conserved; the number of droplets is not. Interfacial tension raises the chemical potential of small droplets (Gibbs–Thomson):

\[ \begin{equation} \mu(R) \approx \mu_{\mathrm{eq}} + \frac{C\,\sigma}{R}, \tag{9}\label{eq:GT} \end{equation} \]

so material flows through the dilute background from small droplets to large ones. This is Ostwald ripening. A dissolved droplet of radius 6 carries about 100 cells of excess A; spread over roughly 12,000 background cells, it raises the background by 0.01–0.02, which is invisible in the dots.

t⟨A⟩⟨P⟩⟨A⟩ where A < 0
200−0.599269495956−7.15×10⁻⁵−0.8723
400−0.599269495956−7.15×10⁻⁵−0.8903
600−0.599269495956−7.15×10⁻⁵−0.8847
800−0.599269495956−7.15×10⁻⁵−0.8781
1000−0.599269495956−7.15×10⁻⁵−0.8799
Table 2. Spatial means from the in-browser solver (α = 0.11, DP = 0.05). The means are fixed to every printed digit; the background level moves as droplets dissolve and are absorbed.

6Extension: A eats P, P regrows

Adding consumption, death, and logistic regrowth gives

\[ \begin{equation} \frac{\partial A}{\partial t} = \nabla^2\Big[ D_A(A^3 - A - \kappa\nabla^2 A) + D_{AP}P \Big] + g\,\tilde A\,P - d\,\tilde A \tag{10}\label{eq:Afeed} \end{equation} \] \[ \begin{equation} \frac{\partial P}{\partial t} = \nabla^2\Big[ D_P P + D_{PA}A \Big] - g\,\tilde A\,P + r\,P\Big(1 - \frac{P}{K}\Big) \tag{11}\label{eq:Pfeed} \end{equation} \]

with \(\tilde A = (1+A)/2\) the local fraction of A material. Mass is no longer conserved. Expected effects: a selected droplet size instead of unlimited coarsening, droplet division when food is plentiful, stronger drift because eating deepens the wake, and possible predator–prey cycles in total A and P. These are expectations, not yet simulated here.

AreaModelRelation to Eqs. 10–11
Tumour growthCahn–Hilliard tumour fraction with a consumed, resupplied nutrient (Hawkins-Daarud, van der Zee & Oden 2012; Garcke, Lam, Sitka & Styles 2016)Closest match: P is the nutrient
Active emulsionsDroplets made from a precursor and degraded by reactions (Zwicker, Hyman, Jülicher et al., 2015–2017)Suppressed ripening, droplet division
Cell coloniesCahn–Hilliard with logistic growth (Khain & Sander 2008)Population dynamics in the A field
Reactive blendsCahn–Hilliard with a linear reaction (Oono & Shiwa; Glotzer, Di Marzio & Muthukumar 1995)Simplest arrested coarsening
EcologyMussel beds as phase separation (Liu, van de Koppel et al., PNAS 2013); Klausmeier vegetation–water; Shigesada–Kawasaki–Teramoto cross-diffusionOrganisms are A, food or water is P
Table 3. Model families where a phase-separating species consumes a regrowing resource.

7Technical notes

The simulation in Figure 1 does not use Crank–Nicolson. It uses a first-order implicit–explicit (IMEX) Euler step with linear stabilization, on a Fourier pseudo-spectral grid. This is the cheapest scheme that handles the stiff fourth-order term and avoids a nonlinear solve.

Space

The box is periodic, \(128\times128\) with \(\Delta x = 1\). Fields are held as Fourier coefficients on wavenumbers \(k = 2\pi m/128\), \(|m| \le 64\), so \(\nabla^2 \to -k^2\) exactly and the largest \(k^2\) is \(2\pi^2 \approx 19.7\). The cubic term is evaluated in real space and transformed back (pseudo-spectral), without dealiasing. The FFT is a hand-written radix-2 Cooley–Tukey transform in double precision; the 2D transform is row transforms, a transpose, row transforms, and a transpose back.

Time stepping

Each right-hand side is split into a stiff linear part, taken at the new time level, and everything else, taken at the old one. With \(N(A) = A^3 - A\) and a stabilization constant \(S\), each Fourier mode advances as

\[ \begin{equation} \big[1 + \Delta t\,k^2(D_A\kappa k^2 + S)\big]\,\hat A^{n+1} = (1 + \Delta t\,S k^2)\,\hat A^{n} - \Delta t\,k^2\big[D_A\,\widehat{N(A^n)} + D_{AP}\,\hat P^{n}\big] \tag{12}\label{eq:stepA} \end{equation} \] \[ \begin{equation} \big(1 + \Delta t\,D_P k^2\big)\,\hat P^{n+1} = \hat P^{n} - \Delta t\,k^2 D_{PA}\,\hat A^{n} \tag{13}\label{eq:stepP} \end{equation} \]

Modes decouple, so the implicit solve is one division per mode. A step costs two FFTs: an inverse transform of \(\hat A\) to form \(N(A)\), and a forward transform of \(N\). Two more inverse transforms per frame produce the fields for drawing.

How each term is treated

Alternatives for the nonlinearity

The standard options, from cheapest to most robust:

For the non-reciprocal system there is no energy to preserve, so the energy-stability arguments behind convex splitting and SAV no longer apply directly. ETDRK4 is the natural upgrade for accuracy.

Conservation

Every term on the right of Eqs. 5–6 carries a factor \(k^2\), so the \(k = 0\) coefficients of \(\hat A\) and \(\hat P\) never change. The means are conserved to round-off (Table 2). A translation test gave the same result to \(1.7\times10^{-14}\) after shifting a state by (40, 70) cells on the torus and evolving it for 500 steps.

Initial state and drawing

The background is \(A\) = the slider value plus noise of amplitude 0.05. Fourteen droplets of radius 6 are placed at random with the profile \(-\tanh\big((r - r_0)/\sqrt2\big)\), the planar equilibrium interface of Eq. 3 for \(\kappa = 1\). \(P\) starts as noise of amplitude 0.05. The page takes 1500 steps before the first frame. Each cell has two fixed random dot sites per species; a site shows a dot when the scaled local density exceeds that site's threshold. \(A\) is scaled from \([-0.9, 0.9]\) to \([0,1]\). \(P\) is scaled by a running maximum of \(|P|\), so blue dot density is relative, not absolute.

QuantityValue
Grid128 × 128, Δx = 1, periodic
Time step Δt0.1
Stabilization S2
DA, κ1, 1
DAP, DPA−α, +α (default α = 0.11)
DP0.05 (default)
Cost per step2 FFTs, about 1.1 ms in a browser
Table 4. Numerical parameters of the simulation in Figure 1.

8Further reading