betaFeedback

Reading v1View latest version

WorksComputer Science

DropFENO: A Data-Free, Physics-Based Operator-Learning Framework for Droplet Equilibria on Mixed-Wettability Patterned Surfaces

Abstract

We study whether the equilibrium of a drop on a chemically patterned wall, and the capillary forces that the pattern exerts on it, can be learned from the free energy alone, without solution data. We introduce DropFENO (Droplet Free-Energy Neural Operator), in which Fourier (FNO), tensor-factorised Fourier (TFNO) and wavelet (WNO) neural operators are trained in two and three dimensions on the diffuse-interface free energy of the drop, with its volume and centroid constraints imposed as penalties. The predictions are tested with the virtual-work (Hellmann–Feynman) theorem: the force exerted by the pattern on a held drop is a wall integral of the equilibrium field which, for a drop straddling a wettability step, reduces identically to the contact-line law F=σΔcos⁡θ LF=σΔ\cosθ\,L. DropFENO reproduces this law and the energy landscape of a drop crossing a step, and reproduces the pinning of the receding contact line less accurately. The residual error lies in the soft modes of the capillary energy (small displacements of the interface and the contact line). This explains why the energy is predicted accurately while contact-line positions are not, and why on striped walls the operators return Cassie-like states instead of the edge-pinned states whose anisotropic apparent angles the reference minimiser shares with lattice Boltzmann simulations. On held-out patterned walls the median energy of the predictions lies \NdEmedRandFnoTwo of the free-drop energy above the best known minimum (FNO, two dimensions), at about one hundredth of the cost of a per-instance minimisation, and a refinement by a classical solver started from the prediction returns the reference state in almost all cases.

main.pdf

Introduction

A drop on a chemically patterned wall settles into a shape that minimises its free energy under the constraint of fixed volume. The minimiser encodes most of what one wants to know about the wetting of the pattern: the apparent contact angles and their anisotropy on stripes [3, 12], the pinned or depinned state of the contact line at a wettability edge [9, 25, 15], and, through the dependence of the minimum energy on the position of the drop, the capillary force that drives a drop across a wettability step or gradient [4, 28]. Computing it is a constrained, nonconvex variational problem that must be solved anew for every pattern and every position. Surface Evolver [2], diffuse-interface energy minimisation and the long-time limit of lattice Boltzmann or phase-field dynamics are the standard tools ([21, 20, 19]); each solve costs from seconds to hours, and parametric studies over pattern families multiply that cost.

Neural operators learn maps between function spaces from samples of their action and evaluate them at a cost independent of the solver that produced the samples ([14] (Lu et al. 2021; Li et al. 2021)). The Fourier neural operator (FNO) of Li et al. (2021), its tensor-factorised variant (TFNO) ([13]) and the wavelet neural operator (WNO) ([29]) are three representative architectures, which parametrise the global kernel respectively by dense spectral weights, by a low-rank canonical-polyadic factorisation of those weights, and by a mixing of wavelet coefficients. Trained on data, they are surrogates of a numerical solver and inherit both its cost (the training set) and its errors. Physics-informed training replaces or supplements the data by the residual of the governing equations ([26, 30] (Li et al. 2024)), and data-driven and physics-informed networks have been applied to the classification of wetting states and to droplet wetting and spreading on rough surfaces ([22, 23 b]); for problems that derive from a variational principle the residual can be replaced by the energy itself, as in the deep Ritz method for single instances ([7]) and in variational operator learning ([10]). For a free-boundary problem with a contact line, the energy formulation is convenient: the Euler–Lagrange equation is a nonlinear elliptic equation with a nonlinear boundary condition at the wall and Lagrange multipliers for the volume and centroid constraints, whereas the energy is a sum of local terms that involve at most first derivatives of the field, and its constrained minimiser satisfies all of these conditions.

Two questions remain open, and they are the subject of this paper; we address them with a framework that we call DropFENO (Droplet Free-Energy Neural Operator). The first is whether an operator trained only on the free energy (with no solution field ever computed, stored or shown to it) reaches the accuracy of a converged per-instance minimiser over a broad family of chemically heterogeneous walls, in two and in three dimensions, and how the three architectures compare when the solution is a sharp interface meeting a patterned wall. The second, which is more important for the physics, is how to verify such a prediction when no data exist. Pointwise agreement with a reference field is necessary but says little about the physics, because the quantities of interest in wetting (forces, angles, pinning) are functionals of the field that can be sensitive to small errors near the contact line. We show that the constrained-minimisation structure itself provides a precise test. By the Hellmann–Feynman (virtual-work) theorem the force exerted by the pattern on a drop held at equilibrium is an explicit wall integral of the equilibrium field, and for a drop straddling a wettability step this integral reduces identically to the contact-line force law F=σΔcos⁡θLF = \sigma\Delta\cos\theta L, with LL the wetted chord on the step. The force law is thus an exact property of the true minimiser which an approximate field does not satisfy in general; testing an operator against it, and against the slope of its own energy landscape, measures whether it has learnt the physics, not only how close it is to a reference solution.

The problem we consider is the static counterpart of the dynamics in our two- and three-dimensional lattice Boltzmann simulations of drops that migrate across hybrid and striped wettability patterns. There, the force that holds a drop on a step was found to follow the contact-line law with the chord of the footprint, and stripes produced a pronounced anisotropy of the apparent angles. Here the same questions are asked of the equilibrium problem, without any dynamics and without any data. We show that (i) DropFENO reproduces the contact-line force law, the energy landscape of a drop crossing a step and, less accurately, the pinning of the rear contact line at the step, as emergent properties of predictions made independently at each position; (ii) all three DropFENO operators, trained only on the energy, predict equilibrium fields whose median energy lies within a fraction of a per cent of the converged minimum, almost always in the basin of the correct state, at a cost about two orders of magnitude lower than the per-instance solve; (iii) their residual error is concentrated in the soft modes of the energy (displacements of the interface and of the contact line), which explains why the energy is accurate while contact-line positions are not, why on striped walls the operators return Cassie-like averaged states instead of the edge-pinned states that set the apparent angles, and why a classical solver started from a prediction converges faster only to a coarse tolerance; and (iv) the architectures differ: the dense FNO is the most accurate, the factorised TFNO comes within a factor of three of it with 150 times fewer parameters, and the local wavelet kernel of the WNO brings no advantage for a problem whose response to the pattern is non-local. A by-product of the analysis is a constraint artefact of general relevance to diffuse-interface and physics-informed calculations: a volume constraint imposed through the linear liquid fraction supersaturates the vapour by the Gibbs–Thomson (Kelvin) effect, which in three dimensions can hold a third of the liquid or evaporate the drop altogether and which biases the force inferred from the energy landscape; an indicator that is flat in the bulk phases removes it exactly.

The paper is organised as follows. § 3 states the variational problem, derives the virtual-work force and the contact-line law, and describes the discretisation. § 4 describes the data-free training, the three operators and the reference solver. § 4.2 verifies the reference minimiser against Young’s law, the vapour state and the force identity, with both choices of the liquid indicator. § 5 assesses the accuracy, basin selection and cost of the operators over the pattern family. § 6 turns to the physics: the force on a drop crossing a step in two and three dimensions, and the apparent angles on stripes. § 6.2 examines the factorisation rank, the prior, the training budget and the transfer to a finer grid; § 6.3 interprets the results through the stiffness of the operator error, and § 6.4 concludes.

Problem formulation

Free energy of a drop on a chemically patterned wall

We consider a liquid drop in contact with its vapour above a flat, chemically heterogeneous wall y=0y=0 (Figure 1). The two fluids are described by a phase field ϕ(x)∈[−1,1]\phi(\mathbf{x}) \in[-1,1] (ϕ=+1\phi=+1 liquid, ϕ=−1\phi=-1 vapour) and the wall by its local Young angle θ(x,z)\theta(x,z). The equilibrium is the minimiser of the Cahn–Hilliard free energy with the wall energy of Cahn [5]; Jacqmin [11]

Section of the computational box and wall pattern with drop

Figure 1. The equilibrium problem. (a) Section of the computational box: a liquid drop (ϕ=+1\phi= +1) in its vapour (ϕ=−1\phi= -1), separated by a diffuse interface of width ∼ϵ\sim\epsilon, rests on a wall whose Young angle varies along the wall (strip: hydrophobic blue, hydrophilic orange; profile below: cos⁡θ(x)\cos\theta(x)). The box is periodic in xx and zz and closed by a no-flux boundary at the top. The centroid constraint holds the drop at X∗X^{*}; its multiplier λx\lambda_x is the external force needed to hold it there, equal and opposite to the capillary force FxF_x exerted by the pattern. (b) The same problem in three dimensions: the wall field cos⁡θ(x,z)\cos\theta(x,z) and the holding position X∗\boldsymbol{X}^{*} are the only inputs; the drop volume V0V_0 is fixed.

E[ϕ]=∫Ωc[W(ϕ)ϵ+ϵ2∣∇ϕ∣2] dV+∫∂Ωwfw(ϕ;θ) dA,W=14(1−ϕ2)2,c=322,(1)E[\phi] = \int_{\Omega} c\left[\frac{W(\phi)}{\epsilon}+\frac{\epsilon}{2}|\nabla\phi|^2\right]\,\mathrm{d}V+\int_{\partial\Omega_w}f_w(\phi;\theta)\,\mathrm{d}A,\qquad W=\frac{1}{4}(1-\phi^2)^2,\qquad c=\frac{3}{2\sqrt{2}}, \tag*{(1)}

with the cubic wall energy

fw(ϕ;θ)=−cos⁡θ g(ϕ),g(ϕ)=14(3ϕ−ϕ3).(2)f_w(\phi;\theta)=-\cos\theta\,g(\phi),\qquad g(\phi)=\frac{1}{4}(3\phi-\phi^3). \tag*{(2)}

The constant cc is chosen so that the planar profile ϕ=tanh⁡(n/2ϵ)\phi=\tanh(n/\sqrt{2}\epsilon) carries the interfacial tension σ=c∫(W/ϵ+ϵϕ′2/2) dn=1\sigma=c\int(W/\epsilon+\epsilon\phi'^2/2)\,\mathrm{d}n=1, which is our unit of energy per unit area. Because g(±1)=±12g(\pm1)=\pm\frac{1}{2}, the wall energy of a wetted minus a dry patch is fw(+1)−fw(−1)=γSL−γSV=−σcos⁡θf_w(+1)-f_w(-1)=\gamma_{SL}-\gamma_{SV}=-\sigma\cos\theta, so that Young’s law cos⁡θY=(γSV−γSL)/σ\cos\theta_Y=(\gamma_{SV}-\gamma_{SL})/\sigma is recovered locally; the cubic form has g′(±1)=0g'(\pm1)=0 and therefore does not alter the bulk values of ϕ\phi at the wall (Jacqmin [11]). The wettability enters the problem only through the wall field cos⁡θ(x,z)\cos\theta(x,z).

The minimisation is subject to two constraints,

V[ϕ]=∫Ωh(ϕ) dV=V0,X[ϕ]=1V0∫Ωh(ϕ) x∥ dV=X∗,h(ϕ)=12+g(ϕ),(3)\mathcal{V}[\phi]=\int_{\Omega}h(\phi)\,\mathrm{d}V=V_0,\qquad X[\phi]=\frac{1}{V_0}\int_{\Omega}h(\phi)\,\mathbf{x}_{\parallel}\,\mathrm{d}V=X^*,\qquad h(\phi)=\frac{1}{2}+g(\phi), \tag*{(3)}

where x∥=(x,z)\mathbf{x}_{\parallel}=(x,z) are the wall-parallel coordinates. The volume constraint fixes the drop size, V0=43πR03V_0=\frac{4}{3}\pi R_0^3 in three dimensions (πR02\pi R_0^2 per unit depth in two). The liquid indicator hh rises from 0 to 1 across the interface and, like the wall function gg, is flat at ϕ=±1\phi=\pm1; the reason for this choice, rather than the linear liquid fraction 12(1+ϕ)\frac{1}{2}(1+\phi), is explained in § 3.1.2. The centroid constraint pins the drop at a prescribed position X∗X^* on the pattern. On a heterogeneous wall a drop held at an arbitrary X∗X^* is not free: the constraint supplies exactly the external force needed to hold it there, and the constrained minimiser ϕ∗(⋅;X∗)\phi^*(\cdot;X^*) is the static shape of a drop held by that force. Varying X∗X^* therefore traces the quasi-static energy landscape E∗(X∗)E^*(X^*) along which a drop migrates, and its gradient is the capillary driving force.

The virtual-work force and the contact-line force law

Translate the wall pattern rigidly by ξ\boldsymbol{\xi} relative to the fluid, cos⁡θ(x∥)→cos⁡θ(x∥−ξ)\cos\theta(\mathbf{x}_{\parallel}) \to\cos\theta(\mathbf{x}_{\parallel}-\boldsymbol{\xi}). The bulk energy and both constraints are invariant under a simultaneous translation of the fluid, so the constrained minimum energy depends only on the relative position of drop and pattern. Since ϕ∗\phi^* is stationary with respect to all admissible variations, the derivative of E∗E^* with respect to ξ\boldsymbol{\xi} is the explicit (partial) derivative alone: this is the Hellmann–Feynman, or virtual-work, theorem (Feynman, [8]):

F≡−∂E∗∂ξdrop=∫∂Ωw∇∥cos⁡θ g(ϕ∗) dA.(4)\mathbf{F} \equiv-\frac{\partial E^*}{\partial\boldsymbol{\xi}_{\mathrm{drop}}} = \int_{\partial\Omega_w} \nabla_{\parallel}\cos\theta\,g(\phi^*)\,\mathrm{d}A. \tag*{(4)}

F\mathbf{F} is the net capillary force exerted by the pattern on the drop (positive towards higher cos⁡θ\cos\theta). It requires only the wall trace of the equilibrium field, involves no derivative of ϕ∗\phi^*, and is exact at the constrained minimum. For a straight step cos⁡θ=cos⁡θ1+Δcos⁡θ H(x−xs)\cos\theta=\cos\theta_1+\Delta\cos\theta\,H(x-x_s) crossed by the drop, ∇∥cos⁡θ=Δcos⁡θ δ(x−xs) x^\nabla_{\parallel}\cos\theta=\Delta\cos\theta\,\delta(x-x_s)\,\hat{\mathbf{x}}, and equation (4) becomes

Fx=σ Δcos⁡θ∫[g(ϕ∗(xs,0,z))+12]dz≡σ Δcos⁡θ Lϕ,(5)F_x=\sigma\,\Delta\cos\theta\int\left[g\bigl(\phi^*(x_s,0,z)\bigr)+\frac{1}{2}\right]\mathrm{d}z\equiv\sigma\,\Delta\cos\theta\,L_\phi, \tag*{(5)}

where the constant 12\frac{1}{2} is the contribution of the dry wall (and, in a periodic box, of the returning step), and LϕL_\phi is the length of the wetted chord of the footprint on the step line, weighted by the diffuse wall trace. For a sharp interface LϕL_\phi is the geometric chord LL and (5) is the classical contact-line force F=σΔcos⁡θ LF=\sigma\Delta\cos\theta\,L of a drop straddling a wettability step (Brochard, [4]; Subramanian et al., [28]); in two dimensions LL is the unit depth. The derivation shows that the law is an identity of the equilibrium problem whenever the step is straddled, independent of the drop shape, and that it fails as soon as a contact line sits on the step itself, where g(ϕ∗)g(\phi^*) takes intermediate values: this is the pinning regime, in which the contact line is held at the step while its local angle sweeps between the two Young angles (Gibbs, [9]; Oliver et al., [25]).

Equation (4) is the central diagnostic of this paper (Figure 2; its discrete form is algorithm 1). An approximate minimiser, in particular a neural operator, satisfies (4) only to the extent that it is a true constrained minimiser; comparing the force computed from (4) with the finite-difference slope of its own energy landscape and with σΔcos⁡θL\sigma\Delta\cos\theta L tests the physics of the prediction, not its similarity to a target field.

The virtual-work test of contact-line physics

Figure 2. The virtual-work test of contact-line physics. (a) A drop held across a straight wettability step: the footprint cuts a chord of length LL on the step line, and the pattern pulls the drop towards the more wettable side with the force Fx=σΔcos⁡θLF_x = \sigma\Delta\cos\theta L. (b) For any candidate equilibrium field the force can be obtained in three independent ways: (i) from the wall trace through the virtual-work integral (4), (ii) from the slope of the energy landscape E∗(X∗)E^*(X^*) traced by independent solves, and (iii) from the geometric chord through the contact-line law (5). The three coincide at a true constrained minimum when the step is straddled, so their mismatch measures the quality of an approximate minimiser without reference data.

The liquid indicator and the Gibbs–Thomson supersaturation

The choice of the indicator in (3) is not innocuous. In the bulk, stationarity of the Lagrangian requires cW′(ϕ)/ϵ=−λVh′(ϕ)cW'(\phi)/\epsilon=-\lambda_V h'(\phi), where λV\lambda_V is the multiplier of the volume constraint, equal to minus the Laplace pressure p=σκp=\sigma\kappa of the drop. With the linear liquid fraction, h=12(1+ϕ)h=\frac{1}{2}(1+\phi), the right-hand side is the constant p/2p/2, and both bulk phases are shifted off ϕ=±1\phi=\pm1; under the bound ∣ϕ∣≤1|\phi|\le1 only the vapour can respond, and with W′(−1+δ)≃2δW'(-1+\delta)\simeq2\delta it is supersaturated by

δ≃ϵσκ4c.(6)\delta\simeq\frac{\epsilon\sigma\kappa}{4c}. \tag*{(6)}

This is the diffuse-interface counterpart of the Kelvin effect. Its consequences grow with the ratio of the box to the drop volume and with ϵ/R0\epsilon/R_0: the liquid stored in the vapour is 12δ\frac{1}{2}\delta per cell, which is a few per cent of V0V_0 in our two-dimensional box but about a third of V0V_0 in the three-dimensional one, where a drop at θY=140∘\theta_Y=140^\circ evaporates completely into the supersaturated vapour (§ 4.2). The centroid then includes a uniform background that does not move with the drop, so that the multiplier of the centroid constraint, λx=∂E∗/∂X∗\lambda_x=\partial E^*/\partial X^*, exceeds the virtual-work force (4) by the factor V0/VdropV_0/V_{\mathrm{drop}}; and the exchange of mass between drop and vapour is a slow, nearly degenerate mode for iterative minimisers. All three effects are artefacts of the constraint, not of the physics of the drop. With h=12+g(ϕ)h=\frac{1}{2}+g(\phi), which satisfies h′(±1)=0h'(\pm1)=0, the bulk equation is satisfied exactly by ϕ=±1\phi=\pm1, the vapour stays at ϕ=−1\phi=-1, the constraints measure the drop alone, and the centroid translates rigidly with it, so that λx=−Fx\lambda_x=-F_x exactly. We use hh throughout and document the linear choice in § 4.2 for comparison.

Discretisation

The fields live on a uniform grid of unit spacing with ϵ=1\epsilon= 1: Nx×NyN_x \times N_y cells in two dimensions and Nx×Ny×NzN_x \times N_y \times N_z in three, periodic in xx and zz, with the wall below the first cell row and a no-flux boundary at the top. The energy (1) is evaluated with forward differences,

E[ϕ]=c∑cells[W(ϕijk)ϵ+ϵ2∑d(Dd+ϕ)ijk2]−∑i,kcos⁡θik g(ϕi0k),(7)E[\phi] = c \sum_{\mathrm{cells}} \left[ \frac{W(\phi_{ijk})}{\epsilon} + \frac{\epsilon}{2}\sum_d \left(D_d^{+}\phi\right)_{ijk}^{2} \right] - \sum_{i,k}\cos\theta_{ik}\,g(\phi_{i0k}), \tag*{(7)}

and the volume and centroid by the corresponding sums. The virtual-work force (4) is evaluated with a central difference of the sampled wall field, Fx=∑i,k12(cos⁡θi+1,k−cos⁡θi−1,k)g(ϕi0k)F_x = \sum_{i,k} \frac{1}{2}\left(\cos\theta_{i+1,k} - \cos\theta_{i-1,k}\right)g(\phi_{i0k}), which telescopes to Δcos⁡θ\Delta\cos\theta across any step. All computations below are carried out with the discrete problem (7); the reference solutions and the neural operators solve exactly the same problem, so that every difference between them is a difference in the quality of the minimisation.

DropFENO: data-free operator learning of capillary equilibria

The solution operator and the training objective

For a fixed drop volume the equilibrium defines a solution operator

G:(cos⁡θ(x,z),X∗)⟼ϕ∗(x,y,z),(8)\mathcal{G}:\left(\cos\theta(x,z),X^*\right)\longmapsto\phi^*(x,y,z), \tag*{(8)}

from the wall wettability field and the holding position to the constrained energy minimiser. We approximate G\mathcal{G} by a neural operator Gw\mathcal{G}_{\mathbf{w}} with weights w\mathbf{w} and train it by minimising the expected value of the physical objective over a distribution P\mathcal{P} of wall patterns and positions,

L(w)=E(a,X∗)∼P[E[ϕw]E0+μ{(V[ϕw]V0−1)2+∣X[ϕw]−X∗R0∣2}],ϕw=Gw(a,X∗),(9)\mathcal{L}(\mathbf{w})=\mathbb{E}_{(a,X^*)\sim\mathcal{P}}\left[\frac{E[\phi_{\mathbf{w}}]}{E_0}+\mu\left\{\left(\frac{V[\phi_{\mathbf{w}}]}{V_0}-1\right)^2+\left|\frac{X[\phi_{\mathbf{w}}]-X^*}{R_0}\right|^2\right\}\right],\qquad\phi_{\mathbf{w}}=\mathcal{G}_{\mathbf{w}}(a,X^*), \tag*{(9)}

with E0=2πR0σE_0=2\pi R_0\sigma (two dimensions) or 4πR02σ4\pi R_0^2\sigma (three dimensions), the energy of a free drop. No equilibrium field enters (9): the inputs a=cos⁡θ(x,z)a=\cos\theta(x,z) are drawn afresh at every iteration from P\mathcal{P} and the only training signal is the energy of the predicted field, which is the same energy that defines the physics. Because the minimiser of (9) over all maps is the pointwise constrained minimiser of EE (for μ→∞\mu\to\infty), training is an amortised optimisation: the cost of solving the variational problem is paid once, during training, for the whole family P\mathcal{P} ([1]). The pipeline is drawn in Figure 3 and listed as algorithm 2. The penalty weight is increased geometrically from μ=10\mu=10 to μ=103\mu=10^3 over the first 60% of training, a continuation that lets the network first learn the shape and then enforce the constraints.

Overall architecture of DropFENO, the data-free operator, and its physics loss

Figure 3. Overall architecture of DropFENO, the data-free operator, and its physics loss. Wall patterns and holding positions are drawn afresh at every iteration from the distribution P\mathcal{P} (Figure 4) and turned into the input channels aa, which include the analytic cap prior ϕ0\phi_0. The neural operator Gw\mathcal{G}_{\boldsymbol{w}} (lifting P\mathcal{P}, four operator blocks, projection QQ; Figure 5 and Figure 6) returns a correction u^w\hat{u}_{\boldsymbol{w}} that is added to u0=artanh⁡ϕ0u_0=\operatorname{artanh}\phi_0 and mapped through tanh⁡\tanh to the field ϕw\phi_{\boldsymbol{w}}, which satisfies ∣ϕw∣<1|\phi_{\boldsymbol{w}}|<1 by construction. The physics loss scores ϕw\phi_{\boldsymbol{w}} by its free energy (7) and by penalties on the volume and centroid constraints, equation (9); its gradient with respect to the weights, obtained by automatic differentiation through the energy, is the only training signal (dashed). No equilibrium field is used at any stage of training. After training, a prediction costs one forward pass and can optionally be refined by the reference minimiser (Figure 8).

The pattern distribution. P\mathcal{P} is a mixture of six families of wall fields (four in two dimensions) (Figure 4), with the two Young angles of each pattern drawn uniformly in [35∘,150∘][35^\circ,150^\circ]: uniform walls; a single smoothed step 12[1+tanh⁡(x/0.75)]\frac{1}{2}[1+\tanh(x/0.75)] placed within ±1.2R0\pm1.2R_0 of the drop; stripes of period P∈[5,32]P\in[5,32], area fraction [0.2,0.8][0.2,0.8] and random phase, running along zz or (in three dimensions) along xx; checkerboards (P∈[8,32](P \in[8,32], three dimensions only); and smooth random patches obtained by Gaussian filtering and thresholding white noise. The drop position is drawn within ±R0/2\pm R_0/2 of the box centre. The distribution is generic: it contains no information on any of the test configurations of § 4.2–§ 6 beyond their pattern family, and the test patterns (e.g. the specific Young angles of the lattice Boltzmann walls used for comparison) are never drawn exactly.

Six families of wall patterns: uniform, step, stripes parallel to z, stripes parallel to x, checkerboard, and random patches

Figure 4. The six families of wall patterns of the training distribution P\mathcal{P} (top view of the wall, schematic). Each training instance draws a family, two Young angles and the parameters of the family, and a holding position within ±R0/2\pm R_0/2 of the box centre (dashed circle: drop footprint of radius R0R_0). The test patterns of § 4.2–§ 6 are never drawn exactly.

Output parametrisation and the analytic prior

The network does not output ϕ\phi directly. It outputs a correction u^\hat{u} to the field u0=artanh⁡ϕ0u_0 = \operatorname{artanh}\phi_0 of a diffuse spherical cap,

ϕw=tanh⁡(u0+u^w),ϕ0=−tanh⁡(∣x−xc∣−Rc2 ϵ)(10)\phi_{\mathrm{w}} = \tanh\left(u_0 + \hat{u}_{\mathrm{w}}\right), \qquad\phi_0 = -\tanh\left(\frac{\lvert\mathbf{x} - \mathbf{x}_{\mathrm{c}}\rvert- R_{\mathrm{c}}}{\sqrt{2}\,\epsilon}\right) \tag*{(10)}

where the cap of volume V0V_0, centred above X∗\mathbf{X}^{*}, has the Cassie-like mean angle cos⁡θˉ=⟨cos⁡θ⟩\cos\bar{\theta} = \langle\cos\theta\rangle over a disc of radius R0R_0 under the drop. The tanh enforces the bound ∣ϕ∣<1\lvert\phi\rvert< 1 and the prior supplies the trivial part of the answer (the size and the mean shape), so that the network learns the response to the pattern; the last layer is initialised to zero, so that training starts from the prior. The prior is itself an approximation built from the inputs alone and contains no solution data; an ablation without it (§ 6.2) quantifies its role.

Architectures

All three operators share the structure of Li et al. (2021): a pointwise lifting PP of the input channels to a width dvd_v, L=4L = 4 blocks

vl+1=GELU⁡(Klvl+Wlvl)(11)v_{l+1} = \operatorname{GELU}\left(\mathcal{K}_l v_l + W_l v_l\right) \tag*{(11)}

with WlW_l a pointwise linear map, and a pointwise projection QQ to the scalar u^\hat{u}. The inputs are the wall field cos⁡θ\cos\theta extended uniformly in yy, the normalised coordinates y/Nyy/N_y, (x−X∗)/R0(x-X^{*})/R_0 and, in three dimensions, (z−Z∗)/R0(z-Z^{*})/R_0, and the prior ϕ0\phi_0. Because yy is not periodic, the latent fields are zero-padded by eight cells at the top before the global operators and cropped afterwards. The three operators differ only in the global kernel Kl\mathcal{K}_l. Figure 5 shows the forward pass with its tensor shapes, Figure 6 the three kernels, and algorithm 3 lists one prediction.

Block diagram of the operator

Figure 5. Block diagram of the operator. (a) Forward pass in two dimensions with the tensor shapes: pointwise lifting PP of the four input channels to the latent width dvd_v, zero padding of the non-periodic yy direction, L=4L = 4 operator blocks, cropping and pointwise projection QQ to the scalar correction u^\hat{u}. In three dimensions the input has five channels and the grid is 64×32×6464 \times32 \times64. (b) One operator block, equation (11): a global kernel Kl\mathcal{K}_l and a pointwise linear map WlW_l act in parallel, followed by a GELU activation. The three operators differ only in Kl\mathcal{K}_l (Figure 6).

The three global kernels

Figure 6. The three global kernels Kl\mathcal{K}_l. (a) FNO: Fourier transform, multiplication of the retained modes by dense complex matrices, inverse transform. (b) TFNO: the same convolution with the weight tensor replaced by its rank-rr canonical-polyadic factorisation, equation (12), drawn as a tensor network (one leg per index) and applied factor by factor. (c) WNO: a three-level Haar analysis of the padded latent field, learned channel mixing of the coarsest approximation and detail coefficients at every coarse location, and the inverse transform with the finer detail bands set to zero. Parameter counts are per block for the two-dimensional configurations of Table 1.

FNO. Kl\mathcal{K}_l is a convolution evaluated in Fourier space, (Kv)(x)=F−1[R(k)Fv(k)](\mathcal{K}v)(\mathbf{x}) = \mathcal{F}^{-1}\left[R(\mathbf{k})\mathcal{F}v(\mathbf{k})\right], with a dense complex weight R∈Cdv×dvR \in\mathbb{C}^{d_v \times d_v} on each retained mode ∣kd∣≤kmax⁡,d\lvert k_d\rvert\le k_{\max,d} and zero elsewhere (Li et al., 2021). Because the field is real, only non-negative wavenumbers are kept in the last direction, so that md=2kmax⁡,dm_d = 2k_{\max,d} modes are retained in every other direction and md=kmax⁡,dm_d = k_{\max,d} in the last one; one kernel has dv2∏dmdd_v^2 \prod_d m_d complex weights.

TFNO. The weight tensor RiokR_{iok} of the FNO is replaced by a complex canonical-polyadic (CP) factorisation of rank rr [13],

Riokxkykz=∑q=1rλqAiqBoqKkxq(x)Kkyq(y)Kkzq(z)(12)R_{iok_xk_yk_z} = \sum_{q=1}^{r} \lambda_q A_{iq} B_{oq} K_{k_xq}^{(x)} K_{k_yq}^{(y)} K_{k_zq}^{(z)} \tag*{(12)}

which reduces the number of complex weights of one kernel from dv2∏dmdd_v^2 \prod_d m_d to r(1+2dv+∑dmd)r(1 + 2d_v + \sum_d m_d) and is applied without forming RR, by contracting the factors in sequence.

WNO. KlK_l acts in a wavelet basis [29]: a J=3J = 3-level orthonormal Haar transform of vlv_l, a learned channel mixing, location by location, of the approximation and the detail coefficients at the coarsest level, and the inverse transform with the finer detail coefficients set to zero. The kernel is local in space and in scale, in contrast to the global Fourier modes of the FNO, which makes it a candidate for a field that contains a thin interface and a pinned contact line.

We refer to the three variants as DropFENO-FNO, DropFENO-TFNO and DropFENO-WNO, and to them collectively as DropFENO; in figures and tables they are labelled FNO, TFNO and WNO. They are implemented in PyTorch from these definitions; Table 1 lists their sizes.

operatorglobal kernel KlK_lmodes or griddvd_vrrparams (2D)params (3D)
FNOdense spectral24×1624 \times16 / 12×8×1212 \times8 \times1232/2032 / 20–6.30 M14.75 M
TFNOCP-factorised spectral24×1624 \times16 / 12×8×1212 \times8 \times1232/2032 / 2032/4832 / 4842.8 k39.7 k
WNOHaar wavelet, J=3J = 316×916 \times9 / 8×5×88 \times5 \times832/2032 / 20–2.37 M4.10 M

Table 1. The three operators. Retained Fourier modes (FNO, TFNO) or coarsest-level wavelet grid (WNO), latent width dvd_v and CP rank rr are given as two-dimensional / three-dimensional values; all operators have L=4L = 4 blocks and the same lifting and projection.

Training

Training uses AdamW (weight decay 10−510^{-5}) with a one-cycle learning-rate schedule (peak 10−310^{-3}), gradient clipping at norm one, batches of eight (two dimensions) or two (three dimensions) independently drawn patterns, 6000 (two dimensions) and 4000 (three dimensions) iterations, i.e. 4.8×1044.8 \times10^4 and 8×1038 \times10^3 distinct problem instances, each seen once. The training histories of the energy and of the constraint residuals are shown in Figure 7. All computations run on CPUs: training on four threads per job (two for the extended-budget runs of § 6.2), reference solves and evaluation on two. The configuration, cost and training history of every run, including the ablations, are given in the Supplementary Information (Table S3, Figure S4 and Figure S3).

Data-free training curves for FNO, TFNO, and WNO in two and three dimensions

Figure 7. Data-free training of the three operators in two (a–c) and three (d–f) dimensions. (a,d) Energy of the predicted fields, averaged over each batch of freshly drawn patterns (running mean); the scatter reflects the variety of the patterns, not the optimisation. (b,e) Volume and (c,f) centroid residuals, enforced only through the penalty whose weight grows from 10 to 10310^3 over the first 60% of training.

runiterationsbatchinstancesseedpriorthreadsparams (k)time (h)
FNO 2D6000848,0000yes46,301.20.90
TFNO 2D (r=32r = 32)6000848,0000yes442.80.67
WNO 2D6000848,0000yes42,369.10.86
FNO 2D, seed 16000848,0001yes46,301.20.84
FNO 2D, no prior6000848,0000no46,301.21.02
FNO 2D, 3× budget180008144,0000yes26,301.24.08
TFNO 2D, 3× budget180008144,0000yes242.83.53
TFNO 2D, r=4r = 46000848,0000yes413.90.67
TFNO 2D, r=8r = 86000848,0000yes418.00.98
TFNO 2D, r=64r = 646000848,0000yes475.80.98
FNO 3D400028,0000yes414,749.61.92
TFNO 3D400028,0000yes439.71.20
WNO 3D400028,0000yes44,100.02.26

Table S3. Table S3: Configuration and cost of every training run. Instances: number of distinct random problems seen (iterations × batch). All runs use a peak learning rate of 10−310^{-3} and the penalty continuation μ:10→103\mu: 10 \to10^{3}. Params in thousands; training wall-clock time on the CPU threads listed.

Training schedules with penalty weight and learning rate plots

Figure S3. Figure S3: Training schedules (FNO runs; identical for all operators). (a) Penalty weight μ\mu. (b) Learning rate of the one-cycle schedule (5% warm-up, cosine decay).

Batch-mean energy training history for FNO two-dimensional ablation runs

(a)

Volume residual training history for FNO two-dimensional ablation runs

(b)

Batch-mean energy training history for TFNO two-dimensional ablation runs

(c)

Volume residual training history for TFNO two-dimensional ablation runs

(d)

Figure S4. Training histories of the two-dimensional ablation runs against the fraction of training completed (running mean over 40 logged iterations). (a,b) FNO: main run, second seed, no prior, three times the budget. (c,d) TFNO: CP ranks 4–64 and three times the budget. (a,c) Batch-mean energy; (b,d) volume residual.

Reference solutions

To judge the operators we solve each test instance individually by a classical method: an augmented-Lagrangian loop for the constraints around a bound-constrained quasi-Newton minimisation (L-BFGS-B) of the discrete energy directly in ϕ∈[−1,1]\phi\in[-1, 1], started from the analytic cap, until the constraint residuals are below 10−710^{-7} and the energy is stationary. Working directly in ϕ\phi with bounds, rather than in the tanh parametrisation that is natural for the network, avoids the vanishing gradients of tanh at the bulk values ϕ=±1\phi= \pm1. Because steps and stripes admit several metastable minima, the reference for those problems is the lowest of three solves started from caps with the Cassie-like mean angle and with the two extreme angles of the pattern. The reference solutions are never used in training. The procedure is shown in Figure 8 and listed as algorithm 4.

Figure 8. The reference minimiser used to judge the operators (§ 4.1.4). An augmented-Lagrangian loop enforces the volume and centroid constraints around a bound-constrained quasi-Newton (L-BFGS-B) minimisation of the discrete energy directly in ϕ∈[−1,1]\phi\in[-1,1]. For steps and stripes the loop is run from three caps and the lowest minimum is kept. Started from an operator prediction instead of a cap, the same loop is the “refinement” of § 5. The converged multipliers are the holding force λx\lambda_x and the Laplace pressure.

Algorithm 1 Virtual-work test of a predicted or computed equilibrium (Figure 2)

Require: step pattern cos⁡θ=cos⁡θ1+Δcos⁡θ H(x−xs)\cos\theta= \cos\theta_1 + \Delta\cos\theta\,H(x-x_s); held positions X1∗<⋯<XM∗X_1^* < \cdots< X_M^*; a map X∗↦ϕ∗X^* \mapsto\phi^* (operator or minimiser)

  1. for m=1,…,Mm = 1,\ldots,M do

  2. ϕm∗←ϕ∗(⋅;Xm∗)\phi_m^* \leftarrow\phi^*(\cdot;X_m^*); Em∗←E[ϕm∗]\mathcal{E}_m^* \leftarrow E[\phi_m^*]

  3. Fm←∑i,k12(cos⁡θi+1,k−cos⁡θi−1,k) g(ϕm,i0k)F_m \leftarrow\sum_{i,k}\frac{1}{2}(\cos\theta_{i+1,k}-\cos\theta_{i-1,k})\,g(\phi_{m,i0k}) ▹ virtual-work force, (4)

  4. Lm←L_m \leftarrow wetted chord of the footprint ϕm∗(xs,0,z)>0\phi_m^*(x_s,0,z)>0 on the step line

  5. end for

  6. Sm←−(Em+1∗−Em−1∗)/(Xm+1∗−Xm−1∗)S_m \leftarrow-(\mathcal{E}_{m+1}^*-\mathcal{E}_{m-1}^*)/(X_{m+1}^*-X_{m-1}^*) ▹ slope of the landscape

  7. return FmF_m, SmS_m and σΔcos⁡θ Lm\sigma\Delta\cos\theta\,L_m; agreement of the three for straddled steps tests the physics of ϕ∗\phi^*

Algorithm 2 Data-free training of the solution operator (Figure 3)

Require: pattern distribution P\mathcal{P}; operator Gw\mathcal{G}_{w}; iterations NN; batch size BB; penalty range μ0=10\mu_0=10, μ1=103\mu_1=10^3; peak learning rate η=10−3\eta=10^{-3}

  1. initialise ww, with the last layer of QQ set to zero so that ϕw=ϕ0\phi_w=\phi_0 at the start

  2. for n=0,…,N−1n=0,\ldots,N-1 do

  3. μ←μ0(μ1/μ0)min⁡(1,n/0.6N)\mu\leftarrow\mu_0(\mu_1/\mu_0)^{\min(1,n/0.6N)} ▹ penalty continuation

  4. for b=1,…,Bb=1,\ldots,B do ▹ fresh problems, never repeated

  5. draw (cos⁡θb(x,z),Xb∗)∼P(\cos\theta_b(x,z),X_b^*)\sim\mathcal{P}

  6. (ab,u0,b)←Inputs⁡(cos⁡θb,Xb∗)(a_b,u_{0,b})\leftarrow\operatorname{Inputs}(\cos\theta_b,X_b^*) ▹ algorithm 3, lines 1–4

  7. ϕb←tanh⁡(u0,b+Gw(ab))\phi_b\leftarrow\tanh(u_{0,b}+\mathcal{G}_{w}(a_b))

  8. end for

  9. L←1B∑b[E[ϕb]E0+μ{(V[ϕb]V0−1)2+∣X[ϕb]−Xb∗∣2R02}]\mathcal{L}\leftarrow\frac{1}{B}\sum_b\left[\frac{E[\phi_b]}{E_0}+\mu\left\{\left(\frac{\mathcal{V}[\phi_b]}{V_0}-1\right)^2+\frac{|X[\phi_b]-X_b^*|^2}{R_0^2}\right\}\right] ▹ equation (9)

  10. g←∇wLg\leftarrow\nabla_w\mathcal{L} by automatic differentiation through EE, V\mathcal{V}, XX; clip ∥g∥≤1\lVert g\rVert\leq1

  11. w←AdamW⁡(w,g)w\leftarrow\operatorname{AdamW}(w,g) with one-cycle learning rate (peak η\eta), weight decay 10−510^{-5}

  12. end for

  13. return ww ▹ no equilibrium field has been used

Algorithm 3 Prediction of an equilibrium field (one forward pass; Figure 5 and Figure 6)

Require: wall field cos⁡θ(x,z)\cos\theta(x,z), holding position X∗X^*, drop volume V0V_0, trained weights ww

  1. cos⁡θˉ←\cos\bar{\theta}\leftarrow mean of cos⁡θ\cos\theta over the disc ∣x∥−X∗∣<R0|x_{\parallel}-X^*|<R_0 ▹ Cassie-like angle

  2. ϕ0←\phi_0\leftarrow diffuse spherical cap of volume V0V_0 and angle θˉ\bar{\theta} above X∗X^*, equation (10)

  3. u0←artanh⁡ϕ0u_0\leftarrow\operatorname{artanh}\phi_0

  4. a←[cos⁡θ, y/Ny, (x−X∗)/R0, (z−Z∗)/R0, ϕ0]a\leftarrow[\cos\theta,\ y/N_y,\ (x-X^*)/R_0,\ (z-Z^*)/R_0,\ \phi_0] ▹ cos⁡θ\cos\theta extended in yy; zz channel in 3D only

  5. v0←P(a)v_0\leftarrow P(a); pad v0v_0 by 8 cells at the top in yy

  6. for l=0,…,L−1l=0,\ldots,L-1 do

  7. vl+1←GELU⁡(Klvl+Wlvl)v_{l+1}\leftarrow\operatorname{GELU}(\mathcal{K}_lv_l+W_lv_l) ▹ FNO, TFNO or WNO kernel Kl\mathcal{K}_l

  8. end for

  9. crop vLv_L in yy; u^←Q(vL)\hat{u}\leftarrow Q(v_L)

  10. return ϕw=tanh⁡(u0+u^)\phi_w=\tanh(u_0+\hat{u})

Algorithm 4 Reference minimiser and refinement (Figure 8)

Require: wall field, X∗X^{*}, V0V_{0}; set of initial fields S\mathcal{S} (three caps for steps and stripes, one cap otherwise, or an operator prediction ϕw\phi_{\mathbf{w}} for refinement)

  1. for each ϕ(0)∈S\phi^{(0)} \in\mathcal{S} do

  2. ϕ←ϕ(0)\phi\leftarrow\phi^{(0)}; λ←0\lambda\leftarrow0; choose μ\mu

  3. repeat

  4. ϕ←arg min⁡−1≤ϕ≤1 E[ϕ]+λ⋅g[ϕ]+μ2∣g[ϕ]∣2\phi\leftarrow\underset{-1 \leq\phi\leq1}{\operatorname{arg\,min}}\ E[\phi] + \lambda\cdot g[\phi] + \frac{\mu}{2}|g[\phi]|^{2} by L-BFGS-B, warm-started at ϕ\phi

  5. λ←λ+μ g[ϕ]\lambda\leftarrow\lambda+ \mu\,g[\phi] \hfill ▹ g=(VV0−1,X−X∗,Z−Z∗)\triangleright\ g = \left(\frac{V}{V_{0}} - 1, X - X^{*}, Z - Z^{*}\right)

  6. until ∣g[ϕ]∣<10−7|g[\phi]| < 10^{-7} and EE stationary

  7. store the candidate (ϕ,E[ϕ],λ)(\phi,E[\phi],\lambda)

  8. end for

  9. return the candidate of lowest energy; its multiplier λx=∂E∗/∂X∗\lambda_x = \partial E^{*}/\partial X^{*} is the holding force

Verification of the reference and of the virtual-work force

Before the operators are assessed, the reference solutions must be shown to be minimisers of the discrete problem (7), and the virtual-work force (4) must be shown to be the physical force. Three independent checks are made; their results are collected in Figure 9. Each check is also run with the linear indicator 12(1+ϕ)\frac{1}{2}(1+\phi), to make concrete the constraint artefact of § 3.1.2.

Three-panel verification plots

Figure 9. Verification of the reference minimiser and of the constraint formulation. (a) Apparent contact angle minus Young angle on uniform walls, for the reference and the operators, in two (open symbols) and three (filled) dimensions. (b) Supersaturation δ=1+ϕv\delta= 1 + \phi_v of the vapour with the linear liquid fraction as indicator, against the Gibbs–Thomson prediction (6); with the indicator hh used in this work δ\delta vanishes to machine precision. (c) Force on a drop held across the 128.8∘→79.7∘128.8^\circ\to79.7^\circ step: virtual-work force (4) and slope of the energy landscape, with the indicator hh (filled) and with the linear fraction (open), for which the slope exceeds the force by V0/VdropV_0/V_{\mathrm{drop}}.

Young’s law. On uniform walls with 40∘≤θY≤150∘40^\circ\le\theta_Y \le150^\circ the apparent contact angle, obtained from a least-squares circle fitted to the ϕ=0\phi=0 contour above the first three cell rows, departs from the prescribed Young angle by at most 2.3∘2.3^\circ in two dimensions (r.m.s. 1.3∘1.3^\circ) and by at most 3.9∘3.9^\circ in the two orthogonal sections of the three-dimensional caps (Figure 9a). The departure is not a convergence error: restarting the minimiser from caps of angle θY±25∘\theta_Y \pm25^\circ returns the same state to 10−9E010^{-9}E_0 (the one exception, a start at 15∘15^\circ on the 40∘40^\circ wall, ends in a distinct state 0.027E00.027E_0 higher). It is a property of the discrete Cahn wall at an interface width of about three cells, and it is the problem the operators are trained to solve. All comparisons below are therefore made with the discrete minimiser, not with the continuum angle.

The vapour and the drop volume. With the indicator hh the bulk vapour of every reference solution sits exactly at the bound ϕ=−1\phi=-1 (median deviation zero in both dimensions; the only departures are the exponential tails of the interface), and the volume constraint measures the drop alone. With the linear indicator it does not (Figure 9b). For the two-dimensional drop at θY=90∘\theta_Y=90^\circ the vapour sits at ϕv=−1+0.0107\phi_v=-1+0.0107, and over the twelve uniform walls the measured shift δ=1+ϕv\delta=1+\phi_v, divided by the Gibbs–Thomson prediction (6) with κ=1/R\kappa=1/R the curvature of the fitted circle, is 1.01 on average (largest deviation 0.09). In three dimensions, where κ=2/R\kappa=2/R and the box is larger relative to the drop, the shift at 90∘90^\circ is 0.0375, a fraction 0.34 of the prescribed volume is held in the vapour, and the drop at 140∘140^\circ evaporates entirely (largest value of ϕ\phi in the box, −0.89-0.89). The supersaturation is quantitatively the Kelvin effect of the diffuse model, and with the linear indicator the ‘drop’ that a solver returns is smaller than the drop that was asked for.

The force on a drop straddling a step. For a drop held across a step, the virtual-work force must equal σΔcos⁡θ\sigma\Delta\cos\theta per unit depth independently of the drop shape and of the held position, as long as the step is straddled (equation (5)). Over the 26 straddling states of the three two-dimensional step walls of Table 2 it does so to within 7×10−47\times10^{-4}. The slope of the energy landscape, −dE∗/dX∗-\mathrm{d}\mathcal{E}^*/\mathrm{d}X^*, a quantity computed independently from the energies alone, is (1.000(1.000–1.001) σΔcos⁡θ1.001)\,\sigma\Delta\cos\theta, and the drop itself, measured by the midpoint of its contact lines, moves by 1.0001.000–1.0021.002 times the displacement of the centroid (Figure 9c, filled symbols). The force, the energy landscape and the kinematics are thus mutually consistent, and the contact-line law F=σΔcos⁡θ LF=\sigma\Delta\cos\theta\,L is recovered as an identity of the minimiser. With the linear indicator the force is unchanged, but the slope is (1.063(1.063–1.070) σΔcos⁡θ1.070)\,\sigma\Delta\cos\theta and the drop moves by 1.0781.078–1.0811.081 times the centroid (open symbols): the centroid includes the vapour excess, which does not move with the drop, and the multiplier of the centroid constraint overestimates the capillary force by the factor V0/VdropV_0/V_{\mathrm{drop}}. The same consideration applies to any diffuse-interface or pseudopotential calculation of the force on a drop from the response of its energy to a displacement of its centre of mass, and to any physics-informed loss built on a linear volume fraction.

setcontent2D3D
uniformθY=40\theta_Y = 40–150150 (2D), 60, 90, 120, 140 (3D)124
step128.8 →\to 79.7, 128.8 →\to 104.5, 143.7 →\to 74.5 (2D; first only in 3D); −1.6≤X∗/R0≤1.6-1.6 \le X^*/R_0 \le1.6519
stripesmixed 143.7 ∣\mid 74.5, f=12f = \frac{1}{2}, half-period w=2w = 2–12; area-fraction 128.8 ∣\mid 79.7, P=6P = 6, 12 (3D: 6), f=16f = \frac{1}{6}–56\frac{5}{6}; sharp edges1611
randomtraining distribution, independent seed10020

Table 2. Held-out test problems. None of them is used in training. The reference minimum of the step and stripe problems is the lowest of three starts; for all problems it is further compared with the refined operator predictions (§ 5). Angles in degrees; R0=16R_0 = 16 (2D) and 12 (3D) lattice units.

Accuracy of the data-free operators

Test problems and error measures

The operators are tested on problems that were never used in training (Table 2): (i) twelve uniform walls with 40∘≤θY≤150∘40^\circ\le\theta_Y \le150^\circ; (ii) a drop held at seventeen positions −1.6≤X∗/R0≤1.6-1.6 \le X^*/R_0 \le1.6 across three sharp-edged steps whose Young angles are those of the lattice Boltzmann walls used for comparison (128.8∘→79.7∘128.8^\circ\to79.7^\circ, 128.8∘→104.5∘128.8^\circ\to104.5^\circ, 143.7∘→74.5∘143.7^\circ\to74.5^\circ); (iii) striped walls with the same angles, of half-period w=2w = 2–1212 and area fraction f=1/6f = 1/6–5/65/6, with sharp (unsmoothed) stripe edges and periods down to P=4P = 4, below the range sampled in training; and (iv) 100 patterns drawn from the training distribution with an independent random seed. The structured sets (ii) and (iii) contain configurations that are deliberately away from the training distribution (held positions up to 1.6R01.6R_0 from the box centre, sharper and narrower stripes), and they are the ones of physical interest.

For each problem we compare the energy of the prediction with the best minimum known for that problem. Because stripes and steps admit several metastable minima, the reference solver is started from three caps (with the Cassie-like mean angle and with the two extreme angles of the pattern) and the lowest minimum is kept; in addition, every operator prediction is used as the starting point of a reference solve (“refined” prediction), and the lowest of all these minima defines EbestE_{\mathrm{best}} (Figure 10). The energy gap (E−Ebest)/E0(E - E_{\mathrm{best}})/E_0 is a strict measure: it is non-negative for any field that satisfies the constraints, it vanishes only at the global minimum among those found, and it does not depend on the choice of norm, which affects pointwise field errors at a diffuse interface. We also report the relative L2L^2 error of the field, the intersection-over-union (IoU) of the liquid regions ϕ>0\phi> 0, the constraint residuals and the apparent angles. Per-run and per-family statistics, the distributions of the constraint residuals and the forces of every random problem are tabulated in the Supplementary Information (Table S4–Table S7, Figure S5 and Figure S6).

Evaluation workflow showing held-out test sets, operator, refinement, reference minimiser, best energy, and prediction metrics

Figure 10. Evaluation workflow. Every held-out problem is predicted by the operator in one forward pass and solved independently by the reference minimiser; the prediction is also refined by the same minimiser. The best known minimum EbestE_{\mathrm{best}} is the lowest of the reference and refined energies, and every prediction is scored against it and against the reference field by the metrics listed.

Table S4: Accuracy of every run on the random held-out set: energy gap (median and 90th percentile), relative field error, intersection-over-union of the liquid regions, volume and centroid residuals, and absolute error of the virtual-work force FxF_x (in units of σ\sigma) with respect to the reference. Gaps of the main runs are taken to the best known minimum EbestE_{\mathrm{best}}, those of the ablation runs to the reference minimum.

runmed. δE\delta EP90 δE\delta E∥δϕ∥\lVert\delta\phi\rVertIoU∣δV∣/V0\lvert\delta V\rvert/V_0∣δX∣/R0\lvert\delta X\rvert/R_0∣δFx∣\lvert\delta F_x\rvert
FNO 2D6.7×10−46.7 \times10^{-4}6.8×10−36.8 \times10^{-3}0.0570.9651.7×10−41.7 \times10^{-4}1.0×10−41.0 \times10^{-4}0.001
TFNO 2D (r=32r = 32)2.0×10−32.0 \times10^{-3}2.2×10−22.2 \times10^{-2}0.1200.9362.5×10−42.5 \times10^{-4}1.2×10−41.2 \times10^{-4}0.002
WNO 2D1.9×10−31.9 \times10^{-3}3.3×10−23.3 \times10^{-2}0.1180.9391.6×10−41.6 \times10^{-4}9.4×10−59.4 \times10^{-5}0.002
FNO 2D, seed 16.2×10−46.2 \times10^{-4}8.1×10−38.1 \times10^{-3}0.0600.9662.0×10−42.0 \times10^{-4}9.6×10−59.6 \times10^{-5}0.001
FNO 2D, no prior3.2×10−33.2 \times10^{-3}1.3×10−21.3 \times10^{-2}0.1380.9292.8×10−42.8 \times10^{-4}2.4×10−42.4 \times10^{-4}0.003
FNO 2D, 3×\times budget1.8×10−41.8 \times10^{-4}2.6×10−32.6 \times10^{-3}0.0260.9862.1×10−42.1 \times10^{-4}6.3×10−56.3 \times10^{-5}0.000
TFNO 2D, 3×\times budget9.0×10−49.0 \times10^{-4}8.0×10−38.0 \times10^{-3}0.0630.9661.9×10−41.9 \times10^{-4}8.6×10−58.6 \times10^{-5}0.001
TFNO 2D, r=4r = 42.3×10−32.3 \times10^{-3}2.4×10−22.4 \times10^{-2}0.1260.9392.6×10−42.6 \times10^{-4}1.4×10−41.4 \times10^{-4}0.002
TFNO 2D, r=8r = 82.0×10−32.0 \times10^{-3}2.3×10−22.3 \times10^{-2}0.1230.9411.9×10−41.9 \times10^{-4}1.0×10−41.0 \times10^{-4}0.002
TFNO 2D, r=64r = 641.7×10−31.7 \times10^{-3}2.2×10−22.2 \times10^{-2}0.1180.9412.3×10−42.3 \times10^{-4}8.3×10−58.3 \times10^{-5}0.001
FNO 3D1.0×10−31.0 \times10^{-3}2.2×10−22.2 \times10^{-2}0.0670.9562.5×10−42.5 \times10^{-4}1.1×10−41.1 \times10^{-4}0.104
TFNO 3D2.0×10−32.0 \times10^{-3}2.9×10−22.9 \times10^{-2}0.0800.9443.7×10−43.7 \times10^{-4}4.7×10−54.7 \times10^{-5}0.075
WNO 3D1.8×10−31.8 \times10^{-3}3.5×10−23.5 \times10^{-2}0.0690.9542.5×10−42.5 \times10^{-4}5.6×10−55.6 \times10^{-5}0.113

Table S4.

operatorsetnniterationstrefinet_{\mathrm{refine}} (s)treft_{\mathrm{ref}} (s)∥ϕrefined−ϕref∥\lVert\phi_{\mathrm{refined}}-\phi_{\mathrm{ref}}\rVert
2D FNOuniform123684.03.74.8×10−84.8 \times10^{-8}
2D FNOstep514786.85.86.0×10−86.0 \times10^{-8}
2D FNOstripes164094.14.37.0×10−87.0 \times10^{-8}
2D FNOrandom1004666.45.86.7×10−86.7 \times10^{-8}
2D TFNOuniform123485.23.72.7×10−82.7 \times10^{-8}
2D TFNOstep514906.85.85.0×10−85.0 \times10^{-8}
2D TFNOstripes164145.64.34.2×10−84.2 \times10^{-8}
2D TFNOrandom1004786.25.86.3×10−86.3 \times10^{-8}
2D WNOuniform123434.73.79.0×10−89.0 \times10^{-8}
2D WNOstep515086.85.84.3×10−84.3 \times10^{-8}
2D WNOstripes164073.94.35.3×10−85.3 \times10^{-8}
2D WNOrandom1004804.75.85.3×10−85.3 \times10^{-8}
3D FNOuniform442887.149.41.0×10−71.0 \times10^{-7}
3D FNOstep9550118.387.17.4×10−87.4 \times10^{-8}
3D TFNOuniform439082.749.49.5×10−89.5 \times10^{-8}
3D TFNOstep9527116.987.17.2×10−87.2 \times10^{-8}
3D WNOuniform441478.849.45.4×10−85.4 \times10^{-8}
3D WNOstep9548104.287.19.2×10−89.2 \times10^{-8}

Table S7. Refinement of the operator predictions by the reference minimiser: number of problems, median number of quasi-Newton iterations and wall-clock time of the refinement, median time of the reference solve from the analytic cap, and median relative field difference between the refined prediction and the reference.

Constraint residual histograms for two- and three-dimensional problems

Figure S5. Constraint residuals of the predictions on all held-out problems: (a,c) volume, (b,d) centroid; (a,b) two dimensions, (c,d) three dimensions. Residuals below 10−710^{-7} are shown at 10−710^{-7}.

Virtual-work force predictions against reference minimiser for two-dimensional and three-dimensional random held-out sets

Figure S6. Virtual-work force FxF_x, equation (4), of the operator predictions against that of the reference minimiser on the random held-out problems: (a) two dimensions, (b) three dimensions. The dotted line is equality.

Energy gap, field error and basin selection

Figure 11 shows predictions for eight held-out two-dimensional problems, and Figure 12 summarises the two-dimensional results. On uniform walls the energy gap of the FNO has a median of 6.0×10−56.0 \times10^{-5} (TFNO 1.9×10−41.9 \times10^{-4}, WNO 3.1×10−43.1 \times10^{-4}), and the apparent angles agree with the reference to a fraction of a degree. On the held-out random patterns the median gaps are 6.7×10−46.7 \times10^{-4}, 2.0×10−32.0 \times10^{-3} and 1.9×10−31.9 \times10^{-3} for the FNO, TFNO and WNO, and the median field errors are 0.057, 0.120 and 0.118. To put the energy gap in perspective: in two dimensions a gap of 10−3E010^{-3}E_0 equals the tension energy of 0.1 lattice units of additional interface, and the median field error of the FNO corresponds to a mean displacement of the ϕ=0\phi= 0 contour by about half a lattice cell. The volume constraint, enforced only through the penalty, is satisfied to a median 1.7×10−41.7 \times10^{-4} (FNO).

Eight held-out two-dimensional problems arranged in four rows and two columns. Legend: FNO interface $\phi=0$ shown as a solid blue line; reference minimiser shown as a dashed black line. Panels (a) uniform: $\delta E=1.0\times10^{-6}$, IoU $=0.996$; (b) uniform: $\delta E=3.1\times10^{-5}$, IoU $=0.998$; (c) step: $\delta E=1.0\times10^{-3}$, IoU $=0.947$; (d) step: $\delta E=5.6\times10^{-3}$, IoU $=0.905$; (e) stripes: $\delta E=1.6\times10^{-3}$, IoU $=0.980$; (f) stripes: $\delta E=4.7\times10^{-5}$, IoU $=0.988$; (g) patches: $\delta E=5.9\times10^{-3}$, IoU $=0.925$; (h) patches: $\delta E=1.5\times10^{-4}$, IoU $=0.989$. The horizontal strip below each drop shows wall wettability $\cos\theta(x)$, ranging from hydrophobic (blue) to hydrophilic (orange).

Figure 11. Held-out two-dimensional problems, two of each pattern family, never seen in training: field predicted by the data-free FNO (shading, with its interface ϕ=0\phi= 0 as a solid line) and interface of the reference minimiser (dashed). The strip below each drop shows the wall wettability cos⁡θ(x)\cos\theta(x), from hydrophobic (blue) to hydrophilic (orange). The energy gap to the best known minimum, δE=(E−Ebest)/E0\delta E = (E - E_{\mathrm{best}})/E_0, and the intersection-over-union of the liquid regions are given in each panel. The same problems predicted by the TFNO and the WNO are shown in the Supplementary Information.

Accuracy and cost in two dimensions, with panels (a) through (d) showing energy gaps, field error, intersection-over-union, and wall-clock time

Figure 12. Accuracy and cost in two dimensions. (a) Energy gap to the best known minimum for every test problem (dots) and its median (bars), by pattern family, for the three operators; structured sets on the left, random held-out patterns on the right. (b) Field error against energy gap. (c) Cumulative distribution of the intersection-over-union of the predicted and reference liquid regions. (d) Wall-clock time per problem on two CPU threads: reference solve from the analytic cap, operator prediction, and reference solve started from the operator prediction. Gaps below 10−610^{-6} are shown at 10−610^{-6}.

The difficulty of a pattern grows with the number of contact-line positions it makes available: the gaps on the uniform walls are an order of magnitude smaller than on steps, stripes and random patches, on which the contact line must be placed at, or pinned on, a wettability edge. The largest gaps (up to 4.8×10−24.8 \times10^{-2}) occur for random patches whose many edges create closely spaced metastable states.

A central question for a data-free operator is whether it selects the right state among several metastable minima. Refining each prediction with the reference solver answers it. On the 100 random problems the refined FNO prediction returns to the reference minimum (same energy to 10−7E010^{-7}E_0, median L2L^2 distance below 10−610^{-6}) in all but 1 case, and in that case it reaches a minimum lower than the best of the three reference starts: the operator, trained to minimise the energy over a whole distribution of problems, found a better state than the classical solver started from the natural initial guesses. The gap on the random patterns is therefore an approximation error inside the correct basin, not a wrong basin. On the structured sets, which lie partly outside the training distribution, the results are less uniform: 1 of the 51 step problems and 2 of the 16 stripe problems are refined into a minimum higher than the reference: the prediction sits in the basin of a neighbouring metastable state, with a different number of stripes wetted or with the front line on the other side of the step, and these cases account for the largest gaps of Figure 12(a). The two other operators behave in the same way: every refined TFNO and WNO prediction on the uniform and random walls returns to the reference minimum (with one exception for the TFNO), and on the structured walls the same few step and stripe problems (1 and 2 for the TFNO, 1 and 2 for the WNO) end in a neighbouring metastable state. Basin selection is therefore a property of the problem (of how close in energy the competing states are) more than of the architecture.

The three architectures. The FNO is the most accurate operator on every family: its median energy gap is three to eight times smaller than that of the TFNO and the WNO on uniform walls, steps and random patterns, but only 1.2–1.6 times smaller on the stripes, where all three operators return similar Cassie-like states (§ 6.1.1; Figure 12a). The TFNO achieves this with 42.8 k parameters, against 6.30 M for the FNO: a factor of about 150 fewer, at the cost of at most a factor of about three in the energy gap. The WNO, whose local wavelet kernel might be expected to suit a thin interface, does not outperform the global Fourier kernels; its errors are largest on the step walls (median 8.0×10−38.0 \times10^{-3}, against 1.0×10−31.0 \times10^{-3} for the FNO), plausibly because there the response of the drop to the pattern is non-local: the position of the far contact line is set by the volume constraint, not by the wall beneath it.

Three dimensions

Table 3 collects the results in both dimensions, and Figure 13 shows the three-dimensional ones by family. In three dimensions the FNO reaches median energy gaps of 6.9×10−56.9 \times10^{-5} on uniform walls, 2.7×10−32.7 \times10^{-3} on stripes and 1.0×10−31.0 \times10^{-3} on the random patterns, comparable to two dimensions, although it was trained on six times fewer problem instances. The step problems are the hardest (8.5×10−38.5 \times10^{-3}, IoU 0.874): the drop is held up to 1.6R01.6R_0 from the step, where in the periodic box it also meets the returning step, and the contact line is pinned on an edge in about half of the held positions (§ 6.1). Refinement of the uniform and step predictions returns the reference minimum in all but 1 case. The ranking of the architectures is that of two dimensions but the spread is smaller: on the random patterns the TFNO reaches 2.0×10−32.0 \times10^{-3} and the WNO 1.8×10−31.8 \times10^{-3}, on the steps 1.2×10−21.2 \times10^{-2} and 1.4×10−21.4 \times10^{-2}, and each of them, after refinement, ends in a higher minimum on one step problem. A three-dimensional prediction costs 628 ms (TFNO 575 ms, WNO 778 ms) against a median of 117.8 s for the reference solve, a factor of about 188.

operatormedian (E−Ebest)/E0(E-E_{\mathrm{best}})/E_0 uniformmedian (E−Ebest)/E0(E-E_{\mathrm{best}})/E_0 stepmedian (E−Ebest)/E0(E-E_{\mathrm{best}})/E_0 stripesmedian (E−Ebest)/E0(E-E_{\mathrm{best}})/E_0 randomIoU randomtopt_{\mathrm{op}} (ms)treft_{\mathrm{ref}} (s)
2D FNO6.0×10−56.0 \times10^{-5}1.0×10−31.0 \times10^{-3}4.2×10−34.2 \times10^{-3}6.7×10−46.7 \times10^{-4}0.965475.8
2D TFNO1.9×10−41.9 \times10^{-4}2.8×10−32.8 \times10^{-3}6.5×10−36.5 \times10^{-3}2.0×10−32.0 \times10^{-3}0.936385.8
2D WNO3.1×10−43.1 \times10^{-4}8.0×10−38.0 \times10^{-3}5.2×10−35.2 \times10^{-3}1.9×10−31.9 \times10^{-3}0.939475.8
3D FNO6.9×10−56.9 \times10^{-5}8.5×10−38.5 \times10^{-3}2.7×10−32.7 \times10^{-3}1.0×10−31.0 \times10^{-3}0.956628117.8
3D TFNO7.7×10−47.7 \times10^{-4}1.2×10−21.2 \times10^{-2}3.4×10−33.4 \times10^{-3}2.0×10−32.0 \times10^{-3}0.944575117.8
3D WNO3.3×10−43.3 \times10^{-4}1.4×10−21.4 \times10^{-2}4.4×10−34.4 \times10^{-3}1.8×10−31.8 \times10^{-3}0.954778117.8

Table 3. Accuracy and cost of the data-free operators on the held-out sets of Table 2: median energy gap (E−Ebest)/E0(E-E_{\mathrm{best}})/E_0 per set, median intersection-over-union of the liquid regions on the random set, wall-clock time of one prediction topt_{\mathrm{op}} and median time of one reference solve treft_{\mathrm{ref}} on the random set. In three dimensions only the uniform and step predictions were refined, so that EbestE_{\mathrm{best}} is the reference minimum for the stripe and random sets.

Accuracy and cost in three dimensions

Figure 13. Accuracy and cost in three dimensions, as in Figure 12. The random held-out set contains twenty patterns, so that some families have a single member. (d) Only the reference solve from the analytic cap and the predictions are timed, since the random three-dimensional problems were not refined. Gaps below 10−610^{-6} are shown at 10−610^{-6}.

Cost

A prediction costs 47 ms (FNO) on two CPU threads, against a median of 5.8 s for one reference solve on the same two threads, a factor of about 123 (Figure 12d); training costs 40–55 minutes per operator in two dimensions and 1.2–2.3 hours in three, on four CPU threads. The training cost of the FNO is recovered after about 570 two-dimensional or 60 three-dimensional problems, and a parametric study such as the force–position curves of § 6.1, which with three starts per position takes the classical solver about a quarter of an hour in two dimensions and forty minutes in three, costs a few seconds. Using the prediction as the initial guess of the reference solver does not shorten the solve, however (median 6.4 s): the multipliers of the augmented Lagrangian start from zero, so that the first subproblem pulls even an accurate initial field away from the constrained minimum, and the quasi-Newton solver retraces most of the path (about 200 iterations in the first outer cycle from either start, for the five random problems that we profiled). Warm-starting the multipliers as well, by fitting them to the predicted field, removes this penalty but helps only at coarse tolerance (§ 6.3). The operator and the classical solver are complementary: the first gives an approximate equilibrium in the correct basin in milliseconds, the second makes it exact when that is needed.

Capillary physics recovered by DropFENO without data

A drop held across a wettability step

Figure 14 follows a two-dimensional drop that is held at successive positions X∗X^* across a step from a hydrophobic (x<0x < 0) to a hydrophilic (x>0x > 0) wall. Each point is an independent constrained minimisation (reference) or an independent forward pass of an operator; the operators were never shown a step at most of these positions relative to the box, nor these Young angles.

Three-panel plots of energy landscape, virtual-work force, and contact-line positions for drops across wettability steps

Figure 14. A two-dimensional drop held at position X∗X^* across a wettability step (step at x=0x = 0, hydrophobic for x<0x < 0): reference minimiser (lines) and operators (symbols), each point an independent solve or prediction. (a) Equilibrium energy landscape for three steps. (b) Virtual-work force (4) in units of σΔcos⁡θ\sigma\Delta\cos\theta; the plateau at unity is the contact-line law (5). (c) Positions of the two contact lines for the 128.8∘→79.7∘128.8^\circ\to79.7^\circ step: the front line snaps onto the hydrophilic side at X∗≃−1.0R0X^* \simeq-1.0R_0, the rear line is pinned at the step for 0.6≲X∗/R0≲1.60.6 \lesssim X^*/R_0 \lesssim1.6.

Energy landscape. The equilibrium energy E∗(X∗)\mathcal{E}^*(X^*) is flat while the drop rests entirely on one wall, falls steeply while it straddles the step, and flattens again once it rests on the hydrophilic wall (Figure 14a). The total decrease of the energy, 0.311E00.311E_0, 0.124E00.124E_0 and 0.398E00.398E_0 for the three steps, is the work that the step can do on a drop that crosses it quasi-statically, and an upper bound on the energy available to a released drop. The operators reproduce the landscape, including the positions of its two kinks, without having been trained on any landscape: each point is a separate prediction, and the landscape emerges from the consistency of the predictions across positions.

Force and the contact-line law. The virtual-work force (Figure 14b) makes the structure explicit. It vanishes while the drop is on one wall, jumps to F=σΔcos⁡θF = \sigma\Delta\cos\theta when the advancing contact line reaches the step (X∗≃−1.0R0X^* \simeq-1.0R_0 for the 128.8∘→79.7∘128.8^\circ\to79.7^\circ step), and keeps exactly this value, independently of the deformation of the drop, for as long as both contact lines are on opposite sides of the step: this is the contact-line force law of equation (5), here an identity of the equilibrium problem. The FNO reproduces the plateau to a median relative error of 3.9×10−63.9 \times10^{-6} in 25 of the 26 straddling states (TFNO 24, WNO 22). The states it misses lie at the snap position itself, where the drop with its front line just before the step and the drop with the front line just past it are nearly degenerate.

Pinning of the receding contact line. Beyond X∗≃0.6R0X^{*} \simeq0.6R_{0} the rear contact line reaches the step and stays there (Figure 14c): instead of crossing onto the hydrophilic wall it is pinned at the edge, while the drop is pulled further. Gibbs’ criterion (Gibbs, [9]; Oliver et al., [25]) allows a contact line on an edge to take any angle between the two Young angles; as X∗X^{*} increases the local angle at the rear decreases from the hydrophobic towards the hydrophilic value, the wall trace g(ϕ∗)g(\phi^{*}) at the step decreases from +12+\frac{1}{2}, and the force falls continuously to zero, reached when the rear line is released at X∗≃1.6R0X^{*} \simeq1.6R_{0}. The response of the drop to the step is therefore strongly asymmetric: the advancing line snaps onto the hydrophilic side, the receding line is pinned at the edge. In the pinning regime the force is a functional of the field at the contact line itself, which is the most difficult output of the problem, and the operators are least accurate there (mean absolute error of F/σΔcos⁡θF/\sigma\Delta\cos\theta of 0.15 for the FNO over the 13 pinned states): their rear contact line reaches the step too late, so that the predicted drop still straddles the step and carries the full force over part of the range in which the reference drop is already pinned.

The three-dimensional step: the chord law. In three dimensions the force on a straddling drop is no longer a constant: by (5) it is σΔcos⁡θ\sigma\Delta\cos\theta times the chord LL that the footprint cuts on the step line, which grows from zero as the advancing line reaches the step, passes through a maximum close to the footprint diameter, and falls again as the rear line approaches (Figure 15). In the periodic box a drop held far from the step also touches the returning step, and the law then reads Fx=σΔcos⁡θ(L−Lret)F_{x}=\sigma\Delta\cos\theta\left(L-L_{\mathrm{ret}}\right). The reference force, computed from the wall trace alone, follows this law with the geometric chords measured on the ϕ=0\phi=0 contour to within 1.0 per cent in all 5 straddling states, with a maximum Fx=1.80 σR0F_{x}=1.80\,\sigma R_{0}. In the remaining states a contact line sits on an edge (the rear line on the step, or the advancing line on the returning, hydrophilic-to-hydrophobic edge), and the force departs from the chord law exactly as the pinning argument of § 3.1.1 predicts: the chord on the line vanishes while g(ϕ∗)g(\phi^{*}) on the edge, and with it the force, takes intermediate values. This is the statement, here derived and verified for the equilibrium problem, that our lattice Boltzmann simulations found to hold for the force that drives a three-dimensional drop across a step. The FNO reproduces the force of the straddling states to a median relative error of 0.04 (TFNO 0.06, WNO 0.07); over all held positions, including the pinned ones, its median error is 0.15 of the largest force (TFNO 0.19, WNO 0.21). As in two dimensions, the operators err where a contact line sits on an edge (Figure 15).

Virtual-work force and chord length for a three-dimensional drop held across a step

Figure 15. A three-dimensional drop (R0=12R_0 = 12, 128.8∘→79.7∘128.8^\circ\to79.7^\circ) held at nine positions across a step in a periodic box. (a) Virtual-work force (4) of the reference minimiser (filled circles) and of the operators (open symbols) against the contact-line law σΔcos⁡θ(L−Lret)\sigma\Delta\cos\theta(L-L_{\mathrm{ret}}) evaluated with the reference chords (line). (b) The chord LL cut by the footprint on the step line and, dashed, the chord LretL_{\mathrm{ret}} on the returning step of the periodic box. Shaded: states in which a contact line sits on an edge, where the force departs from the chord law.

Apparent angles on striped walls

Two dimensions: pinning on the stripe edges. On a striped wall the equilibrium drop has its contact lines on stripe edges whenever it can, and its apparent angle is then not the Cassie–Baxter angle but whatever angle the volume and the positions of the two pinned lines impose (Brandon et al., [3]; Kusumaatmaja & Yeomans, [15]). The reference minimisers show this clearly (Figure 16a,Figure 16b): for the 143.7° | 74.5° stripes of equal width the apparent angle varies non-monotonically between 74° and 124° as the half-period ww grows from 2 to 12 cells, about the Cassie–Baxter value of 105.6°; on the 128.8° | 79.7° walls of varying area fraction it stays within 9° of the Cassie–Baxter curve for the shorter period P=6P=6 but departs from it by more than 20° for P=12P=12, for which the drop spans only about two periods. The operators do not resolve this selection. Their apparent angles stay within 101°–107° (FNO) on the equal-width stripes, close to the Cassie–Baxter angle, with an r.m.s. deviation from the reference of 13° (TFNO 16∘16^\circ, WNO 16∘16^\circ) over all stripe problems: the predicted contact lines sit between edges rather than on them, and the predicted drop is an average over the pattern. Because a contact-line displacement along a stripe is one of the softest modes of the energy (§ 6.3), these predictions are nevertheless close in energy to the pinned state (median gap 4.2×10−3E04.2 \times10^{-3}E_0), and refinement carries all but 2 of them into the reference minimum. The apparent angle on a striped wall is thus the quantity most sensitive to the soft-mode error of an energy-trained operator, and the one for which a refinement step is indispensable.

Apparent contact angles on striped walls

Figure 16. Apparent contact angles on striped walls. (a) Two dimensions, mixed 143.7∘∣74.5∘143.7^\circ|74.5^\circ stripes of width ww; the Cassie–Baxter angle and the two Young angles are shown as horizontal lines. (b) Two dimensions, 128.8∘∣79.7∘128.8^\circ|79.7^\circ stripes of period PP and hydrophilic fraction ff, with the Cassie–Baxter angle. (c) Three dimensions, mixed stripes: angle across (θ⊥\theta_\perp, filled) and along (θ∥\theta_\parallel, open) the stripes, compared with the drops of our three-dimensional lattice Boltzmann simulations at the same w/R0w/R_0.

Three dimensions: the anisotropy of the apparent angle. On a three-dimensional striped wall the pinning acts only across the stripes: the contact line can slide along a stripe edge but not across it, so that the footprint is shaped by the stripes (Figure 17k). The apparent angle measured in the section across the stripes, θ⊥\theta_\perp, and along them, θ∥\theta_\parallel, then differ. The reference minimisers on the 143.7∘∣74.5∘143.7^\circ|74.5^\circ stripes show an anisotropy of up to 36∘36^\circ (Figure 16c): for w/R0=1w/R_0 = 1, θ⊥=130∘\theta_\perp= 130^\circ and θ∥=94∘\theta_\parallel= 94^\circ; for w/R0=1/2w/R_0 = 1/2, θ⊥=93∘\theta_\perp= 93^\circ and θ∥=103∘\theta_\parallel= 103^\circ; for w/R0=1/3w/R_0 = 1/3, θ⊥=116∘\theta_\perp= 116^\circ and θ∥=104∘\theta_\parallel= 104^\circ. These are, to within six degrees, the angles of the drops in our three-dimensional lattice Boltzmann simulations at comparable w/R0w/R_0 (131∘131^\circ and 89∘89^\circ at w/R0=1w/R_0 = 1, 96∘96^\circ and 100∘100^\circ at 1/21/2, 122∘122^\circ and 105∘105^\circ at 0.30.3), obtained there at a 2.5 times larger drop radius and as the end state of an impact and relaxation. The agreement shows that the anisotropy found in the dynamic simulations is an equilibrium property, set by which stripe edges the contact line is pinned on, and that the non-monotonic dependence on w/R0w/R_0 reflects the discrete number of stripes spanned by the footprint. The operators, in contrast, predict nearly isotropic drops: θ⊥\theta_\perp and θ∥\theta_\parallel differ by at most 1∘1^\circ for the FNO (TFNO 1∘1^\circ, WNO 0∘0^\circ) and stay near the Cassie–Baxter value, as in two dimensions, while the energy gap on these walls remains small (median 2.7×10−3E02.7 \times10^{-3}E_0). The anisotropy of the apparent angle is, with the pinned receding line on a step, the clearest example of a physically important quantity that lies in the soft modes of the energy, and therefore of a quantity that an energy-trained operator resolves least accurately.

Held-out three-dimensional problems for uniform, step, stripes and random patterned walls

Figure 17. Held-out three-dimensional problems (R0=12R_0 = 12, 64×32×6464 \times32 \times64 cells): a uniform wall at θY=120∘\theta_Y = 120^\circ, a drop held centrally on the 128.8∘→79.7∘128.8^\circ\to79.7^\circ step, 143.7∘∣74.5∘143.7^\circ\mid74.5^\circ stripes of half-period w=4w = 4 and a checkerboard. (a–d) Drop surface ϕ=0\phi= 0 predicted by the data-free FNO above the patterned wall (colour: cos⁡θ\cos\theta, blue hydrophobic, orange hydrophilic), shown over the central 44×4444 \times44 cells. (e–h) Section through the centre of the drop perpendicular to zz: predicted field (shading), its interface (solid) and the interface of the reference minimiser (dashed). (i–l) Wetted footprint ϕ(x,0,z)=0\phi(x, 0, z) = 0 of the FNO (solid) and of the reference (dashed) over the wall pattern.

Ablations: factorisation, prior, budget and resolution

Rank of the factorised kernel. Figure 18(a) places every two-dimensional operator in the plane of parameter count and median energy gap on the random held-out set. The CP rank of the TFNO matters little: rank 4 (13.9 k parameters) gives 2.3×10−32.3 \times10^{-3}, rank 8 gives 2.0×10−32.0 \times10^{-3}, rank 32 (the default, 42.8 k) 2.0×10−32.0 \times10^{-3} and rank 64 (75.8 k) 1.7×10−31.7 \times10^{-3}, against 6.7×10−46.7 \times10^{-4} for the dense FNO with 6.30 M parameters. A sixteen-fold increase of the rank thus improves the gap by less than a factor of 1.4, and the remaining factor of about 2.5 to the dense FNO is not closed within the range of ranks tested: the dense spectral weights of the FNO are compressible to below one per cent of their size at a moderate, not a negligible, loss of accuracy.

Ablations in two dimensions: three-panel plots comparing trainable parameters, cap-prior patterns, and zero-shot transfer

Figure 18. Ablations in two dimensions. (a) Median energy gap on the random held-out set against the number of trainable parameters: TFNO for CP ranks r=4r = 4–6464, FNO, WNO, a second FNO seed and FNO and TFNO trained with three times the budget. (b) FNO with and without the analytic cap prior, by pattern family. (c) Zero-shot transfer to a grid refined by a factor of two (same physical problem, ϵ\epsilon and R0R_0 doubled in lattice units).

Seed and training budget. A second FNO trained from a different initialisation and with a different stream of training patterns reaches 6.2×10−46.2 \times10^{-4} (first seed 6.7×10−46.7 \times10^{-4}), and its median gaps on the individual pattern families differ from those of the first by at most about 20%, except on the uniform members of the set, where the gaps are smallest and differ by 40%. With this single pair of seeds as an indication of the spread, the differences between architectures reported in § 5, which are factors of two or more, are significant, whereas differences of a few tens of per cent are not. Tripling the number of training iterations, and with it the number of distinct training patterns, reduces the median gap of the FNO from 6.7×10−46.7 \times10^{-4} to 1.8×10−41.8 \times10^{-4}, a factor of nearly four, and that of the TFNO from 2.0×10−32.0 \times10^{-3} to 9.0×10−49.0 \times10^{-4} (Figure 18a). The operators are therefore far from saturated at the budget used in the rest of the paper, whose accuracies are conservative; in the data-free setting additional training instances cost only the compute to draw and evaluate them. Even with three times the budget, however, the TFNO does not reach the FNO trained with the base budget, so that the advantage of the dense kernel is not an artefact of under-training.

The analytic prior. Without the cap prior of equation (10) (the network then outputs ϕ=tanh⁡u^\phi= \tanh\hat{u} directly, with the same inputs other than ϕ0\phi_0), the FNO still learns the constrained equilibrium, but its median gap on the random set rises from 6.7×10−46.7 \times10^{-4} to 3.2×10−33.2 \times10^{-3}, and on every pattern family it is two to four times larger, on uniform walls more than twenty times (Figure 18b). The prior is not required, but it is worth more than tripling the training budget: it supplies the drop size and the mean shape, which the network otherwise has to learn from the volume and centroid penalties alone, and leaves the network to learn only the response of the drop to the pattern.

Resolution. Neural operators are often credited with discretisation invariance. For the present problem the claim fails, for a physical reason. Refining the grid by a factor of two at fixed physical problem (drop radius and interface width both doubled in lattice units, step width scaled accordingly) and evaluating the trained operators without retraining increases the median energy gap on uniform walls from 2.5×10−52.5 \times10^{-5} to 2.7×10−32.7 \times10^{-3} for the FNO and on the step problems from 1.0×10−31.0 \times10^{-3} to 9.2×10−39.2 \times10^{-3} (Figure 18c). The discrete energy depends on the interface width in lattice units, which sets both the contact angle actually realised by the Cahn wall (§ 4.2) and the cost of the interface, and an operator trained at ϵ=1\epsilon= 1 has learnt the minimiser of that discrete problem, not of its continuum limit. Retraining at the finer resolution costs no data, only compute.

Discussion

Where the error of an energy-trained operator lives. An operator trained on the energy is penalised for an error e=ϕw−ϕ∗e = \phi_{\mathbf{w}} - \phi^{*} in proportion to 12e⋅He\frac{1}{2} e \cdot H e, with HH the Hessian of the energy at the minimiser, and not in proportion to ∣e∣2\lvert e\rvert^{2}. It is therefore free to leave error in the soft directions of HH and is driven hard to remove it from the stiff ones. The Rayleigh quotient e⋅He/e⋅ee \cdot H e/e \cdot e of the actual FNO error on the random patterns has a median of 0.09 (TFNO 0.23, WNO 0.26), against 0.56 for smooth random perturbations of the same interfacial support and 3.80 for white ones (Figure 19a); the translation mode of the interface, the softest physically meaningful direction, has 0.031. The residual error is thus a small displacement or tilt of the interface and of the contact line, not a distortion of the interface profile. Three observations of § 5–§ 6 follow from this. First, the energy gap is small while the field error is not, because the second is dominated by modes that cost almost no energy. Second, the quantities that are first order in these soft modes (the position of a contact line on an edge, and with it the force in the pinning regime) are the least accurate outputs, whereas the force on a straddling drop, which depends only on which side of the step each line lies, is reproduced to a few parts per million. Third, the prediction accelerates a classical solver only for a coarse tolerance. In a single stiff-penalty quasi-Newton pass with multipliers fitted to the initial field, the FNO prediction reaches a gap of 10−3E010^{-3} E_{0} after a median of 25 iterations, against 313 from the analytic cap (median ratio 6.0 over 30 random problems); but to reach 10−4E010^{-4} E_{0} and 10−5E010^{-5} E_{0} it still needs 363 and 598 iterations, a gain of only 1.8 and 1.5, because the remaining error lies precisely in the ill-conditioned directions that limit the convergence of the solver (Figure 19b). The operator and the solver are, in this sense, complementary rather than redundant.

Where the error of an energy-trained operator lives: stiffness distribution and energy gap along a quasi-Newton solve

Figure 19. Where the error of an energy-trained operator lives. (a) Stiffness e⋅He/e⋅ee \cdot He/e \cdot e of the operator error e=ϕw−ϕ∗e = \phi_{\mathbf{w}} - \phi^* on the 100 random held-out problems, with HH the Hessian of the discrete energy at the reference minimiser, against interface-localised smooth and white noise and the translation mode of the interface (dotted). (b) Energy gap along a quasi-Newton solve started from the analytic cap and from the FNO prediction (median and quartiles over 30 random problems); dotted lines mark the tolerances of § 6.3.

The virtual-work test as a verification principle. Without data, the usual measures of a surrogate (agreement with a held-out solution) presuppose the classical solver that the surrogate is meant to replace. The virtual-work identity ((4)) supplies a test that does not: an operator that has learnt the constrained minimiser must return a force that equals the slope of its own energy landscape and, on a straddled step, the contact-line value σΔcos⁡θL\sigma\Delta\cos\theta L. Both checks require only forward passes of the operator. The same construction applies to any operator trained on a variational principle with a parametric external field (the Hellmann–Feynman force on a nucleus in electronic-structure surrogates, the configurational force on a defect in elasticity, the force on an inclusion in a membrane), and it tests the physics of a prediction rather than its similarity to a reference field.

Constraints in physics-informed losses. The Gibbs–Thomson artefact of § 3.1.2 is not specific to neural operators. It affects any diffuse-interface calculation that imposes the volume or the centre of mass through the linear liquid fraction (a common choice in phase-field energy minimisation and in physics-informed losses for two-phase problems), and its size grows with the ratio of the box to the drop volume and with the interface width relative to the drop: the vapour is shifted by δ≃ϵσκ/4c\delta\simeq\epsilon\sigma\kappa/4c, and each unit volume of vapour holds a liquid volume δ/2\delta/2. In our three-dimensional box it holds a third of the liquid in the vapour and evaporates a drop at 140°. An indicator hh with h′(±1)=0h'(\pm1) = 0 adds no cost and removes it exactly; we recommend it for any such loss.

Limitations. The diffuse interface is three cells wide and the drops are 16 (two dimensions) and 12 (three dimensions) cells in radius. The operators solve the discrete problem at this resolution, including its lattice-induced departures from Young’s law, and do not transfer zero-shot to a finer grid (§ 6.2). The problem is static: the operators predict equilibria and the landscape E∗(X∗)\mathcal{E}^{*}(X^{*}), not the dissipative dynamics along it, and the hysteresis of a moving contact line enters only through the multiplicity of metastable states. Which of these states an operator returns is decided by its training and is not controlled; on structured walls a few predictions lie in the basin of a neighbouring state. Finally, the pattern distribution is generic but finite; patterns far outside it, such as stripes much narrower than the interface, are outside the scope of the present operators.

Conclusions

DropFENO, a set of neural operators trained on the free energy alone without any solution field, learns the capillary equilibrium of drops on chemically patterned walls in two and three dimensions to within a small fraction of the free-drop energy, at a cost two orders of magnitude below that of a per-instance minimisation, and almost always in the basin of the reference state. The constrained-minimisation structure of the problem supplies the means to verify them without data. The virtual-work theorem makes the capillary force an explicit functional of the equilibrium field; on a straddled step it reduces identically to the contact-line law F=σΔcos⁡θLF = \sigma\Delta\cos\theta L, which the reference minimiser satisfies to better than one part in a thousand in two dimensions and to about one per cent with the geometric chord in three, and which the operators reproduce as an emergent property of predictions made independently at each position. The comparison identifies the limits of energy-based training: the error lives in the soft modes of the energy, so that contact-line positions on edges (the pinned receding line at a step, the edge-pinned lines that set the apparent angle on stripes) are the least accurate outputs, while the energy and the force on a straddling drop are the most accurate. For these outputs a short refinement by a classical solver, started from the prediction, recovers the pinned state in nearly all cases. These accuracies are conservative: tripling the training budget, which without data costs only compute, reduces the median gap of the FNO by a further factor of nearly four. Among the architectures the dense FNO is the most accurate, the CP-factorised TFNO trades a factor of three in accuracy for a factor of 150 in size, and the local wavelet kernel of the WNO brings no advantage. Finally, the choice of the liquid indicator in the constraints matters: the linear liquid fraction creates a Gibbs–Thomson supersaturation of the vapour that can hold a third of the liquid and biases forces inferred from energy landscapes, and an indicator flat in the bulk phases removes it. Both the virtual-work test and the constraint remedy carry over to any operator or physics-informed network trained on a variational principle with constraints.

Supplementary Information

The Supplementary Information contains a notation table, the configuration and cost of every training run, per-run, per-family and per-angle accuracy tables, refinement statistics, training histories and schedules, constraint residuals, virtual-work forces of every random problem, galleries of the TFNO and WNO predictions, schematics of the wavelet decomposition and of the input layout, and the pattern sampler and stiffness diagnostic as algorithms.

Data and code availability

The DropFENO source code (the free energy and constraints, the FNO, TFNO and WNO operators, the data-free training, the reference minimiser, and the scripts that reproduce every figure and table) is available at https://github.com/Ganesh1591/DropFENO and will be made public upon publication. No training data exist: the operators are trained on the free energy alone.

Author contributions (CRediT)

G.S. Meshram: conceptualisation, methodology, software, validation, formal analysis, investigation, visualisation, writing (original draft). P.P. Chakrabarti: supervision, methodology, writing (review and editing). S. Chakraborty: supervision, conceptualisation, writing (review and editing).

Acknowledgements

The authors acknowledge the Computational Facility of the Indian Institute of Technology Kharagpur for providing the computing resources used in this work.

Declaration of interests

The authors report no conflict of interest.

supplementary.pdf

Notation

Table S1 lists the symbols used in the main paper and in this document.

symbolmeaningsymbolmeaning
ϕ\phiphase field, +1+1 liquid, −1-1 vapourG\mathcal{G}, Gw\mathcal{G}_{\mathbf{w}}solution operator, its approximation
θ(x,z)\theta(x,z)local Young angle of the wallw\mathbf{w}trainable weights
ϵ\epsilon, σ\sigmainterface width, tensionaainput channels
WW, gg, hhdouble well, wall and indicator functionsϕ0\phi_0, u0u_0analytic cap prior, artanh⁡ϕ0\operatorname{artanh}\phi_0
E\mathcal{E}, EEcontinuous, discrete free energyu^\hat{u}network output (correction)
VV, V0V_0liquid volume, its targetPP, QQlifting and projection
X\mathbf{X}, X∗\mathbf{X}^{*}drop centroid, holding positionKl\mathcal{K}_l, WlW_lglobal kernel, pointwise map of block ll
R0R_0radius of the free dropdvd_v, LLlatent width, number of blocks
E0E_0free-drop energy (2πR0,4πR02)(2\pi R_0, 4\pi R_0^2)kmax⁡k_{\max}retained Fourier modes
λV\lambda_V, λx\lambda_xmultipliers (pressure, holding force)rrCP rank of the TFNO
FxF_xvirtual-work forceJJHaar levels of the WNO
LL (chord)wetted chord on the step lineμ\mupenalty weight
θapp\theta_{\mathrm{app}}apparent angle (circle fit)δE\delta E(E−Ebest)/E0(E-E_{\mathrm{best}})/E_0
PPpattern distributionIoUintersection-over-union of ϕ>0\phi>0

Table S1. Notation. Lattice units throughout: unit cell spacing, ϵ=1\epsilon=1, σ=1\sigma=1.

Architecture details

Table S2 lists the grids and hyperparameters of the two- and three-dimensional configurations; Figure S1 shows the input layout and the padding, and Figure S2 the wavelet bands used by the WNO kernel (complementing Figure 6 of the main paper). Algorithm S1 lists the pattern sampler of Figure 4 of the main paper.

two dimensionsthree dimensions
grid Nx×NyN_x \times N_y (×Nz)(\times N_z)128×64128 \times6464×32×6464 \times32 \times64
drop radius R0R_0, interface width ϵ\epsilon16,116, 112,112, 1
input channels4455
latent width dvd_v, blocks LL32,432, 420,420, 4
padding of yy inside the operator8 cells8 cells
FNO/TFNO retained modes kmax⁡k_{\max}24×1624 \times1612×8×1212 \times8 \times12
TFNO CP rank rr32324848
WNO Haar levels JJ, coarsest grid3,16×93, 16 \times93,8×5×83, 8 \times5 \times8
lifting PP1×11\times1 conv. to dvd_v, GELU, 1×11\times1 conv.1×11\times1 conv. to dvd_v, GELU, 1×11\times1 conv.
projection QQ1×11\times1 conv. to 4dv4d_v, GELU, 1×11\times1 conv. to 11×11\times1 conv. to 4dv4d_v, GELU, 1×11\times1 conv. to 1
parameters FNO / TFNO / WNO6.30 M / 42.8 k / 2.37 M14.75 M / 39.7 k / 4.10 M

Table S2. Grids and hyperparameters of the operators. Parameter counts are real numbers (complex weights count twice).

Input channels and padding inside the operator

Figure S1. (a) Input channels of the two-dimensional operator: the wall field extended uniformly in yy, the normalised height, the horizontal distance from the holding position and the analytic prior. (b) The latent field is zero-padded by eight cells at the top before the global kernels, because the yy direction is not periodic, and cropped after the last block; the wall lies below the first row.

Haar bands of the padded two-dimensional latent field

Figure S2. Haar bands of the padded two-dimensional latent field after J=3J=3 levels. The WNO kernel mixes the channels of the coarsest approximation a3a_3 and of the three coarsest detail bands d3d_3 with weights that depend on the location on the 16×916 \times9 coarse grid; the finer details are set to zero, so that the kernel is a smooth, local, multi-scale operator and local detail is carried by the pointwise path WlW_l.

Algorithm S1 Pattern sampler P\mathcal{P} (Figure 4 of the main paper)

Require: grid NxN_x (×Nz\times N_z), drop radius R0R_0, random generator

  1. family ←\leftarrow uniform choice among {uniform, step, stripes ∥z\parallel z, patches} (2D) or these with stripes ∥x\parallel x and checkerboard (3D)

  2. θ1,θ2∼U[35∘,150∘]\theta_1,\theta_2 \sim U[35^\circ,150^\circ]; c1=cos⁡θ1c_1=\cos\theta_1, c2=cos⁡θ2c_2=\cos\theta_2

  3. if uniform then f←0f\leftarrow0

  4. else if step then f←12[1+tanh⁡((x−xs)/0.75)]f\leftarrow\frac{1}{2}[1+\tanh((x-x_s)/0.75)]

  5. else if stripes then draw period P∈[5,32]P\in[5,32], fraction ∈[0.2,0.8]\in[0.2,0.8], phase; f←f\leftarrow smoothed periodic indicator

  6. else if checkerboard then draw P∈[8,32]P\in[8,32] and phases; f←12[1+sgn⁡sin⁡2πxPsgn⁡sin⁡2πzP]f\leftarrow\frac{1}{2}[1+\operatorname{sgn}\sin\frac{2\pi x}{P}\operatorname{sgn}\sin\frac{2\pi z}{P}]

  7. else f←f\leftarrow smoothed step of Gaussian-filtered white noise (filter scale k∈[3,10]k\in[3,10]) ▹\triangleright patches

  8. end if

  9. cos⁡θ←c1+(c2−c1)f\cos\theta\leftarrow c_1+(c_2-c_1)f

  10. X∗,Z∗∼U[−R0/2,R0/2]X^*,Z^*\sim U[-R_0/2,R_0/2]; for a step, shift the pattern so that the step lies within ±1.2R0\pm1.2R_0 of X∗X^*

  11. return (cos⁡θ,X∗)(\cos\theta,X^*)

Training runs

Table S3 lists every training run of the study, with its cost. All runs use the same schedules (Figure S3): a geometric continuation of the penalty weight from 10 to 10310^{3} over the first 60% of the iterations and a one-cycle learning rate. Figure S4 compares the histories of the ablation runs of § 6.2 of the main paper. Without the prior the network starts from ϕ≈0\phi\approx0 everywhere, with an energy fourteen times that of the free drop and a volume error of order one, and recovers a drop within the first tenth of training; its volume residual stays about twice as large as that of the runs with the prior until the end. The CP rank has no visible effect on the histories, consistent with panel (a) of Figure 18 of the main paper.

Additional accuracy statistics

Table S4 extends Table 3 of the main paper to every run and adds the 90th percentile of the energy gap, the constraint residuals and the error of the virtual-work force. Table S5 breaks the random held-out set down by family, Table S6 gives the apparent angles on uniform walls behind panel (a) of Figure 9 of the main paper, and Table S7 the cost of the refinement. Figure S5 shows that the penalty enforces the volume and the centroid to a few parts in 10410^{4} for every operator and both dimensions, so that the energy gaps of the main paper are not an artefact of violated constraints. Figure S6 compares the virtual-work force of each prediction with that of the reference on all random problems, and Figure S7 the cost of a refinement started from a prediction with that of a solve started from the analytic cap.

Table S5: Random held-out set broken down by pattern family: number of problems, median energy gap δE=(E−Ebest)/E0\delta E = (E - E_{\mathrm{best}})/E_0 and median IoU of each operator.

familynnFNO δE\delta EFNO IoUTFNO δE\delta ETFNO IoUWNO δE\delta EWNO IoU
2D uniform214.4×10−54.4 \times10^{-5}0.9881.6×10−41.6 \times10^{-4}0.9833.0×10−43.0 \times10^{-4}0.985
2D step261.1×10−31.1 \times10^{-3}0.9543.3×10−33.3 \times10^{-3}0.9155.5×10−35.5 \times10^{-3}0.909
2D stripes ∥z\parallel z201.0×10−31.0 \times10^{-3}0.9702.7×10−32.7 \times10^{-3}0.9442.8×10−32.8 \times10^{-3}0.949
2D patches331.7×10−31.7 \times10^{-3}0.9425.3×10−35.3 \times10^{-3}0.9108.5×10−38.5 \times10^{-3}0.910
3D uniform41.3×10−41.3 \times10^{-4}0.9867.4×10−47.4 \times10^{-4}0.9853.4×10−43.4 \times10^{-4}0.985
3D step32.1×10−42.1 \times10^{-4}0.9841.1×10−31.1 \times10^{-3}0.9755.3×10−45.3 \times10^{-4}0.981
3D stripes ∥z\parallel z48.8×10−38.8 \times10^{-3}0.8651.3×10−21.3 \times10^{-2}0.8571.4×10−21.4 \times10^{-2}0.860
3D stripes ∥x\parallel x12.2×10−22.2 \times10^{-2}0.7522.8×10−22.8 \times10^{-2}0.7413.5×10−23.5 \times10^{-2}0.746
3D patches41.1×10−21.1 \times10^{-2}0.8731.3×10−21.3 \times10^{-2}0.8711.4×10−21.4 \times10^{-2}0.869
3D checkerboard45.7×10−35.7 \times10^{-3}0.9038.1×10−38.1 \times10^{-3}0.8991.1×10−21.1 \times10^{-2}0.898

Table S5.

Table S6: Uniform walls in two dimensions: apparent contact angle (degrees) measured by a circle fit to the interface of the reference minimiser and of each operator, and the energy gap δE\delta E of each operator.

θY\theta_Yref.FNOTFNOWNOδE\delta E FNOδE\delta E TFNOδE\delta E WNO
4037.740.442.840.76.1×10−46.1 \times10^{-4}1.5×10−31.5 \times10^{-3}1.1×10−31.1 \times10^{-3}
5048.450.452.850.53.5×10−43.5 \times10^{-4}9.7×10−49.7 \times10^{-4}7.1×10−47.1 \times10^{-4}
6059.460.262.260.47.0×10−57.0 \times10^{-5}5.6×10−45.6 \times10^{-4}3.1×10−43.1 \times10^{-4}
7068.370.071.470.21.3×10−41.3 \times10^{-4}3.3×10−43.3 \times10^{-4}3.0×10−43.0 \times10^{-4}
8079.879.980.780.02.5×10−52.5 \times10^{-5}1.4×10−41.4 \times10^{-4}1.9×10−41.9 \times10^{-4}
9090.090.190.289.91.0×10−61.0 \times10^{-6}3.7×10−53.7 \times10^{-5}1.9×10−41.9 \times10^{-4}
100101.4100.199.899.89.9×10−59.9 \times10^{-5}2.0×10−42.0 \times10^{-4}3.4×10−43.4 \times10^{-4}
110109.2109.7109.5109.61.0×10−61.0 \times10^{-6}5.5×10−65.5 \times10^{-6}2.0×10−42.0 \times10^{-4}
120118.4119.8119.4119.54.9×10−54.9 \times10^{-5}5.3×10−55.3 \times10^{-5}2.3×10−42.3 \times10^{-4}
130130.6130.1129.4129.41.0×10−61.0 \times10^{-6}1.8×10−41.8 \times10^{-4}2.9×10−42.9 \times10^{-4}
140141.5140.3139.3139.31.0×10−61.0 \times10^{-6}1.6×10−41.6 \times10^{-4}4.6×10−44.6 \times10^{-4}
150150.6150.4149.1149.11.3×10−41.3 \times10^{-4}9.4×10−49.4 \times10^{-4}5.9×10−45.9 \times10^{-4}

Table S6.

Refinement of two-dimensional predictions on the random held-out set

Figure S7. Refinement of the two-dimensional predictions on the random held-out set. (a) Quasi-Newton iterations of the refinement against the energy gap of the prediction. (b) Cumulative distribution of the wall-clock time of the reference solve from the analytic cap and of the refinement of each operator’s prediction.

Galleries of the TFNO and WNO predictions

Figures S8–S11 repeat Figure 11 of the main paper and Figure 17 of the main paper for the other two operators on the same held-out problems.

Eight-panel comparison of TFNO and reference minimiser interfaces for uniform, step, stripes, and patches wall wettability patterns. Legend: TFNO interface $\phi=0$; reference minimiser. Panels (a) uniform: $\delta E=1.6 \times 10^{-4}$, IoU $=0.993$; (b) uniform: $\delta E=1.4 \times 10^{-4}$, IoU $=0.990$; (c) step: $\delta E=3.1 \times 10^{-3}$, IoU $=0.923$; (d) step: $\delta E=1.3 \times 10^{-2}$, IoU $=0.842$; (e) stripes: $\delta E=7.0 \times 10^{-3}$, IoU $=0.915$; (f) stripes: $\delta E=1.0 \times 10^{-6}$, IoU $=0.985$; (g) patches: $\delta E=1.1 \times 10^{-2}$, IoU $=0.879$; (h) patches: $\delta E=1.2 \times 10^{-3}$, IoU $=0.984$. Axes show $x$ and $y$, with wall wettability $\cos\theta(x)$ indicated by the strip below each drop, ranging from $-0.8$ to $0.8.

Figure S8. As figure 11 of the main paper, for the TFNO.

Eight-panel comparison of WNO interface $\phi=0$ and reference minimiser for uniform, step, stripes, and patches wall wettability profiles. Panels (a) uniform: $\delta E=4.1\times10^{-4}$, IoU $=0.993$; (b) uniform: $\delta E=2.2\times10^{-4}$, IoU $=0.994$; (c) step: $\delta E=5.5\times10^{-3}$, IoU $=0.914$; (d) step: $\delta E=3.3\times10^{-2}$, IoU $=0.810$; (e) stripes: $\delta E=7.2\times10^{-3}$, IoU $=0.929$; (f) stripes: $\delta E=2.4\times10^{-4}$, IoU $=0.983$; (g) patches: $\delta E=1.7\times10^{-2}$, IoU $=0.854$; (h) patches: $\delta E=6.0\times10^{-4}$, IoU $=0.986$. Axes show $x$ and $y$; legend: WNO interface $\phi=0$, reference minimiser; colour scale: wall wettability $\cos\theta(x)$ (strip below each drop).

Figure S9. As Figure 11 of the main paper, for the WNO.

Comparison of TFNO interfaces and reference minimisers for uniform, step, stripes, and random patterns

Figure S10. As figure 17 of the main paper, for the TFNO.

Panels (a) uniform, (b) step, (c) stripes, and (d) random; panels (e)–(h) with axes $x$ and $y$; panels (i)–(l) with axes $x$ and $z$. Legend: WNO interface $\phi=0$; reference minimiser.

Figure S11. As figure 17 of the main paper, for the WNO.

Stiffness diagnostic

The diagnostic of panel (a) of Figure 19 of the main paper measures how stiff the error of a prediction is with respect to the energy: the Rayleigh quotient of the error with the Hessian of the discrete energy at the reference minimum, compared with that of interface-localised noise and of the translation mode of the interface (algorithm S2). The Hessian is never formed: each quotient needs one Hessian-vector product by double automatic differentiation of the energy.

Algorithm S2 Stiffness of the operator error (panel (a) of Figure 19 of the main paper)

Require: reference minimum ϕref\phi_{\mathrm{ref}}, prediction ϕw\phi_w, wall field

  1. function Ray(vv)

  2. h←∇ϕ(∇ϕE[ϕref]⋅v)\mathbf{h} \leftarrow\nabla_{\phi}\left(\nabla_{\phi} E[\phi_{\mathrm{ref}}] \cdot v\right) ⊲ Hessian-vector product by autograd

  3. return (v⋅h)/(v⋅v)(v \cdot\mathbf{h})/(v \cdot v)

  4. end function

  5. w←1−ϕref2w \leftarrow1-\phi_{\mathrm{ref}}^2 ⊲ weight localised on the interface

  6. sop←Ray⁡(ϕw−ϕref)s_{\mathrm{op}} \leftarrow\operatorname{Ray}(\phi_w-\phi_{\mathrm{ref}})

  7. swhite←Ray⁡(wξ)s_{\mathrm{white}} \leftarrow\operatorname{Ray}(w \xi), ξ\xi white Gaussian noise

  8. ssmooth←Ray⁡(wξˉ)s_{\mathrm{smooth}} \leftarrow\operatorname{Ray}(w \bar{\xi}), ξˉ\bar{\xi} noise averaged over 7×77 \times7 cells

  9. stransl←Ray⁡(ϕref(x+1)−ϕref(x−1))s_{\mathrm{transl}} \leftarrow\operatorname{Ray}(\phi_{\mathrm{ref}}(x+1)-\phi_{\mathrm{ref}}(x-1)) ⊲ rigid shift of the interface

  10. return sop,swhite,ssmooth,stransls_{\mathrm{op}},s_{\mathrm{white}},s_{\mathrm{smooth}},s_{\mathrm{transl}}

References

  1. [1]Amos, B. 2023 Tutorial on amortized optimization. Found. Trends Mach. Learn. 16, 592–732.
  2. [2]Brakke, K.A. 1992 The Surface Evolver. Exp. Math. 1, 141–165.DOI
  3. [3]Brandon, S., Haimovich, N., Yeger, E. & Marmur, A. 2003 Partial wetting of chemically patterned surfaces: the effect of drop size. J. Colloid Interface Sci. 263, 237–243.pubmed.ncbi.nlm.nih.gov/12804908
  4. [4]Brochard, F. 1989 Motions of droplets on solid surfaces induced by chemical or thermal gradients. Langmuir 5, 432–438.DOI
  5. [5]Cahn, J.W. 1977 Critical point wetting. J. Chem. Phys. 66, 3667–3672.
  6. [7]E, W. & Yu, B. 2018 The deep Ritz method: a deep learning-based numerical algorithm for solving variational problems. Commun. Math. Stat. 6, 1–12.arxiv.org/abs/1710.00211
  7. [8]Feynman, R.P. 1939 Forces in molecules. Phys. Rev. 56, 340–343.DOI
  8. [9]Gibbs, J.W. 1906 The Scientific Papers of J. Willard Gibbs, vol. 1, p. 326. Longmans, Green and Co.
  9. [10]Goswami, S., Yin, M., Yu, Y. & Karniadakis, G.E. 2022 A physics-informed variational DeepONet for predicting crack path in quasi-brittle materials. Comput. Methods Appl. Mech. Engng 391, 114587.DOI
  10. [11]Jacqmin, D. 2000 Contact-line dynamics of a diffuse fluid interface. J. Fluid Mech. 402, 57–88.DOI
  11. [12]Jansen, H.P., Bliznyuk, O., Kooij, E.S., Poelsema, B. & Zandvliet, H.J.W. 2012 Simulating anisotropic droplet shapes on chemically striped patterned surfaces. Langmuir 28, 499–505.DOI
  12. [13]Kossaifi, J., Kovachki, N., Azizzadenesheli, K. & Anandkumar, A. 2023 Multi-grid tensorized Fourier neural operator for high-resolution PDEs. arXiv:2310.00120.arxiv.org/abs/2310.00120
  13. [14]Kovachki, N., Li, Z., Liu, B., Azizzadenesheli, K., Bhattacharya, K., Stuart, A. & Anandkumar, A. 2023 Neural operator: learning maps between function spaces with applications to PDEs. J. Mach. Learn. Res. 24 (89), 1–97.
  14. [15]Kusumaatmaja, H. & Yeomans, J.M. 2007 Modeling contact angle hysteresis on chemically patterned and superhydrophobic surfaces. Langmuir 23, 6019–6032.pubmed.ncbi.nlm.nih.gov/17451253
  15. [19]Meshram, G.S. 2025 Quantifying surface wettability in textured surfaces using two-dimensional pseudopotential multiphase lattice Boltzmann model. J. Adhes. Sci. Technol. 39, 1472–1496.DOI
  16. [20]Meshram, G.S. & Biswal, G. 2024 Effect of solid-fluid interaction strength on water droplet placed centrally at the curved surfaces using lattice Boltzmann method. AIP Conf. Proc. 3108, 030012.DOI
  17. [21]Meshram, G.S., Chakraborty, S., Chakrabarti, P.P., Biswal, G. & Kondaraju, S. 2023 Hydrophobicity analysis of the surfaces with rectangular micropillars using pseudopotential multiphase lattice Boltzmann method. In Conference on Fluid Mechanics and Fluid Power (FMFP 2023), pp. 361–374.
  18. [22]Meshram, G.S., Chakraborty, S. & Chakrabarti, P.P. 2025 Demarcating wetting states classification of the textured surfaces using machine learning. Procedia Comput. Sci. 258, 3672–3681.DOI
  19. [23]Meshram, G.S., Chakrabarti, P.P. & Chakraborty, S. 2026a Learning droplet wetting from kinetic theory: lattice-Boltzmann-driven physics-informed neural networks for rough surfaces. Langmuir.DOI
  20. [25]Oliver, J.F., Huh, C. & Mason, S.G. 1977 Resistance to spreading of liquids by sharp edges. J. Colloid Interface Sci. 59, 568–581.DOI
  21. [26]Raissi, M., Perdikaris, P. & Karniadakis, G.E. 2019 Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys. 378, 686–707.DOI
  22. [28]Subramanian, R.S., Moumen, N. & McLaughlin, J.B. 2005 Motion of a drop on a solid surface due to a wettability gradient. Langmuir 21, 11844–11849.pubmed.ncbi.nlm.nih.gov/16316123
  23. [29]Tripura, T. & Chakraborty, S. 2023 Wavelet neural operator for solving parametric partial differential equations in computational mechanics problems. Comput. Methods Appl. Mech. Engrg 404, 115783.DOI
  24. [30]Wang, S., Wang, H. & Perdikaris, P. 2021 Learning the solution operator of parametric partial differential equations with physics-informed DeepONets. Sci. Adv. 7, eabi8605.

Paper details

vlnc:W_g3eewucd
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents