betaFeedback

WorksPhysics and Astronomy

Dealloying by peritectic melting

Abstract

Peritectic melting of Ti–Ag has been shown experimentally to form bicontinuous structures, but the mechanism remains unclear. Here we use phase-field simulations to show that these structures arise from a morphological instability of liquid film migration in three dimensions: a Ti-rich solid growing through an Ag-rich liquid film develops a branched seaweed or dendritic structure whose side branches coalesce to form handles, generating a high-genus bicontinuous topology. A sharp-interface theory predicts a solidification-front velocity and an initial ligament width that are constant in time, in contrast to liquid metal dealloying; subsequent t1/3t^{1/3} coarsening reproduces the experimentally observed final ligament width.

Impact statement

Computational modeling reveals the interfacial pattern formation mechanism that transforms a homogeneous peritectic alloy into a two-phase bicontinuous structure by unstable liquid film migration and interface coalescence during the melting process.

Introduction

Dealloying is a powerful route for fabricating bicontinuous metals with large interfacial area and useful functional and mechanical properties. These structures enable applications in energy storage [1, 2], catalysis [3, 4], sensing [5, 6], and nanomechanics [7, 8]. Conventional electrochemical dealloying (ECD) selectively dissolves less noble elements from a precursor alloy into an electrolyte [9, 10, 11, 12], while liquid metal dealloying (LMD) uses a molten metal to dissolve the miscible elements and extends bicontinuous structure formation to systems such as Ta–Ti, Fe–Cr–Ni, and Si–Mg [13, 14, 15, 16, 17, 18, 19, 20, 21, 22]. Vapor phase dealloying (VPD) further exploits vapor pressure differences to selectively remove elements into a vacuum [23, 24, 25]. In all these cases, selective removal or dissolution into an external medium drives the formation of an interconnected solid/liquid or solid/vapor morphology.

Recent experiments have shown that isothermal melting of a peritectic Ti50Ag50\mathrm{Ti}_{50}\mathrm{Ag}_{50} parent β\beta alloy can also generate bicontinuous microstructures [26, 27]. This process differs fundamentally from ECD, LMD, and VPD because it does not require an external dealloying medium. Upon heating above the peritectic temperature TpT_p, the metastable β\beta phase decomposes into a Ti-rich α\alpha solid and an Ag-rich liquid. Redistribution of Ti and Ag is therefore supported internally by diffusion within the newly formed liquid phase. Among conventional dealloying routes, LMD provides the closest comparison to peritectic melting (PM) because both processes involve two solid phases and a liquid metal. However, in LMD the liquid is an external reservoir, and solute must diffuse through the growing dealloyed layer to the surrounding melt. This produces time-dependent dealloying kinetics, with the dealloying front velocity v∝1/tv \propto1/\sqrt{t} and the ligament width at the front λ0∼1/v\lambda_0 \sim1/\sqrt{v} decreasing and increasing in time, respectively [15, 20]. In PM, by contrast, the liquid forms locally at the melting front, suggesting that the dealloying kinetics may differ from LMD. However, the mechanism by which PM produces a bicontinuous structure and the scaling laws governing vv and λ0\lambda_0 as a function of time and driving force (superheating ΔT=T−Tp\Delta T = T - T_p) remain unknown.

PM has traditionally been associated with liquid film migration (LFM), in which a thin Ag-rich liquid film advances into the parent β\beta alloy while the α\alpha–liquid and β\beta–liquid interfaces migrate in concert [28, 29]. Viewed as a smooth, uniformly advancing front, LFM is topology-preserving and should not by itself generate a high-genus bicontinuous structure. However, recent electron microscopy observations of Ti–Ag peritectic melting suggest that the bicontinuous α/L\alpha/L network develops through LFM [27]: an α\alpha nucleus formed at a premelted β\beta–β\beta grain boundary remains separated from the neighboring β\beta grain by an Ag-rich liquid film and propagates with the receding β\beta–liquid front. This raises the central question: how can LFM, apparently a planar film-migration process, generate high-genus topology?

Here we use phase-field (PF) simulations of PM to answer this question. To isolate the LFM mechanism from the α\alpha–β\beta–liquid triple-junction geometry, we initiate growth from a Ti-rich α\alpha nucleus fully immersed in the Ag-rich liquid film between neighboring β\beta grains. This avoids imposing uncertain contact angles determined by the interfacial free energies γlα\gamma_{l\alpha}, γlβ\gamma_{l\beta}, and γαβ\gamma_{\alpha\beta}, which are not accurately known for Ti–Ag, while preserving the essential feature of LFM: two solid–liquid interfaces migrating together across a thin liquid layer. Seeding growth heterogeneously from an α\alpha nucleus on a β\beta grain, as observed experimentally [26, 27], instead of homogeneously inside the inter-grain liquid film yields the same large-scale morphological development of the α\alpha–liquid interface. We therefore focus on the homogeneous case, which isolates the mechanism of bicontinuous structure formation from the details of initial nucleus spreading that depend on the triple-junction geometry.

Since crystalline anisotropy is well-known to influence solidification morphologies [30, 31], we consider both cases where γlα\gamma_{l\alpha} is isotropic and weakly anisotropic. The simulations reveal that, in both two and three dimensions (2D and 3D), the α\alpha–liquid interface becomes morphologically unstable forming branched seaweed or dendritic structures with the topology evolving in 3D by coalescence of neighboring branches that form handles and produce a high-genus α/L\alpha/L network. They further show that, in contrast to LMD, vv and λ0\lambda_0 are constant in time, and the structure coarsens behind the front with λ∼t1/3\lambda\sim t^{1/3} as in Ostwald ripening [32, 33] instead of a smaller coarsening exponent varying between 1/4 and 1/3 due to a mix of surface and bulk diffusion in LMD [12, 20]. We extend existing sharp-interface theories of dendritic solidification [30, 34] and LFM [29, 35, 36] to derive the scaling laws v∼ΔT2v \sim\Delta T^2 and λ0∼1/ΔT\lambda_0 \sim1/\Delta T that are validated by simulations. These results identify branching and coalescence within morphologically unstable LFM as the mechanism of bicontinuous structure formation, establishing PM as a dealloying pathway distinct from LMD.

Methods

We simulated peritectic melting of Ti–Ag using a multiPF model adapted from a quantitative formulation for eutectic solidification [37]. The Ag-rich liquid, Ti-rich α\alpha solid, and parent β\beta alloy are represented by phase fields ϕl\phi_l, ϕα\phi_\alpha, and ϕβ\phi_\beta, constrained by ϕl+ϕα+ϕβ=1\phi_l + \phi_\alpha+ \phi_\beta= 1, together with a conserved Ag concentration field cc. The free-energy functional contains gradient, triple-well, and concentration-coupling terms calibrated to the linearized Ti–Ag phase diagram and to the interfacial free energies γlα\gamma_{l\alpha}, γlβ\gamma_{l\beta}, and γαβ\gamma_{\alpha\beta}. The phase fields evolve by dissipative relaxation subject to the PF constraint, while cc evolves by diffusion with mobility restricted to the liquid phase, so that solute transport vanishes in the solids and reduces to Fickian diffusion in the bulk liquid.

To examine the effect of crystalline anisotropy, the α\alpha–liquid interfacial free energy was taken to be either isotropic or weakly anisotropic with cubic symmetry [30, 31]. The equations were solved on a uniform Cartesian grid using second-order central differences and explicit time integration in a CUDA C/C++ implementation. Topological analysis of the dealloyed region was performed by reconstructing the α\alpha–liquid interface using marching cubes [38]; the genus gg and ligament size λ\lambda were then obtained from the triangulated surface and structure-factor analysis. Full model equations, numerical parameters, calibration details, anisotropy implementation, and topology-analysis procedures are provided in Supplementary Notes 1–3, with parameters listed in Supplementary Table 1 [39, 40, 41, 42].

Liquid film migration instability during Ti–Ag peritectic melting

Figure 1. Liquid film migration instability during Ti–Ag peritectic melting. A Ti-rich α\alpha nucleus is placed in an Ag-rich liquid film between parent β\beta grains at ΔT=100\Delta T = 100 K above the peritectic temperature TpT_p. (a–c) In 2D with isotropic α\alpha–liquid interfacial energy, the advancing interface destabilizes into a seaweed morphology. (d) Fourfold anisotropy selects dendritic branches. (e–g) In 3D, side-branch coalescence converts the unstable front into a bicontinuous α/L\alpha/L network. (h) Fourfold anisotropy biases growth along cubic axes while preserving bicontinuity.

Results and Discussion

Morphological instability of liquid film migration and bicontinuous structure formation

As observed experimentally, melting preceding α\alpha nucleation is typically initiated at grain boundaries [26, 27], consistent with the theoretical expectation that the formation of a liquid film becomes energetically favored at a grain boundary when its excess free-energy γgb>2γlβ\gamma_{gb} > 2\gamma_{l\beta}. For this reason, melting is only initiated at a subset of grain boundaries of the initial polycrystalline β\beta microstructure. Since we are interested in elucidating the topology-generating pattern formation mechanism associated with LFM, we do not model the details of the liquid film formation and α\alpha nucleation process. We simply start the simulations from a small initially 2D-circular or 3D-spherical nucleus immersed in a straight liquid film between two β\beta grains. The evolution of such a seed nucleus is shown in Figure 1 in 2D (a-d) and 3D (e-h). In both 2D and 3D, the advancing α\alpha–liquid interface destabilizes while remaining separated from the parent β\beta alloy by a continuous liquid film, forming a branched seaweed when γlα\gamma_{l\alpha} is isotropic and a dendritic morphology when γlα\gamma_{l\alpha} is anisotropic.

Image not extracted

Figure 1 (PDF p. 4).

These distinct branched morphologies are qualitatively expected from previous studies of single-phase dendritic growth that have highlighted the critical role of crystalline anisotropy in enabling dendritic growth along preferred crystal axes [30, 31], and the formation of seaweed structures in the absence of anisotropy [43, 44]. However, this role is only secondary here to the formation mechanism of bicontinuous structures, which can be generated in 3D by both seaweed and dendritic growth as illustrated in Figure 1g,h. More important is the topological transformation mediated by the coalescence of neighboring branches that is further illustrated in Figure 2a,Figure 2b. On the time scale of the simulations of Figure 1d and h, coalescence is only present in the 3D geometry of experimental relevance. This difference can be attributed to the existence of two principal radii of curvature of the α\alpha–liquid boundary in 3D, which facilitates transport of excess Ag in the liquid film, enabling α\alpha–liquid interfaces to approach closely enough to coalesce in 3D but not in 2D. Each coalescence event closes a loop in the α\alpha scaffold and creates a topological handle. Such side-branch coalescence is a known pathway for microstructural connectivity during classical dendritic solidification [45], but here it occurs within a migrating liquid film and generates bicontinuity during PM. Repetition of this elementary operation across the dealloying front converts a collection of dendritic branches into a connected bicontinuous α\alpha/liquid network (Figure 1g). Anisotropy (ϵ1=0.1\epsilon_1 = 0.1, Figure 1h) biases growth along the cubic axes but does not suppress branch coalescence, so that the α\alpha scaffold remains topologically interconnected.

Bicontinuous network formation by side-branch coalescence

Figure 2. Bicontinuous network formation by side-branch coalescence. 3D phase-field simulations at ΔT=100\Delta T = 100 K. (a,b) In a single-nucleus simulation, neighboring α\alpha side branches expand laterally and coalesce, closing loops and creating topological handles. (c,d) In a two-nucleus simulation, separate dendritic envelopes coalesce internally and then impinge, producing a polycrystalline bicontinuous network.

The same mechanism also explains how polycrystalline bicontinuous structures can form. When two α\alpha nuclei with distinct orientations are introduced in the same liquid film, each develops its own dendritic envelope. Branch coalescence first creates connectivity within each envelope, and subsequent impingement produces a polycrystalline bicontinuous network (Figure 2c,Figure 2d). In this case, the grain scale is set primarily by the initial spacing of α\alpha nuclei, while the ligament scale is set by branching and coarsening of the migrating α\alpha–liquid interface. PM therefore naturally separates grain-size selection from ligament-size selection, a distinction that may be important for designing polycrystalline bicontinuous microstructures.

Topology and coarsening

To confirm bicontinuity quantitatively, we computed the genus gg of the α\alpha–liquid interface in cubic sub-volumes of the dealloyed region (Figure 3a). Genus counts independent handles of a closed surface: g=0g = 0 corresponds to a sphere-like topology, g=1g = 1 to a torus, and g≫1g \gg1 to a highly connected network. Details of the marching-cubes reconstruction and genus calculation are given in Supplementary Note 3.

Figure 3. Topology and coarsening of the bicontinuous network. (a) Sub-volumes used to measure the genus gg and ligament size λ\lambda. (b) Evolution of g−1g - 1, confirming a high-genus topology; the inset shows the final morphology. (c) Evolution of the cubed ligament size λ3\lambda^3, with curves shifted in time to align local coarsening ages. The inset corresponds to the cube regions marked by diagonal segments 1–3 in panel (a). (d) The scaled genus (g−1)λ3/V(g - 1)\lambda^3/V, where VV is the analyzed volume, approaches ∼2\sim2.

The genus increases rapidly during the active branching and coalescence stage, reaching values far above unity in all analyzed regions (Figure 3b). This confirms that the α\alpha scaffold is not merely branched, but topologically bicontinuous. At later times, gg decreases gradually as capillarity-driven coarsening removes small ligaments and eliminates some handles, while remaining well above unity. Thus, the network retains bicontinuity even as its characteristic length scale increases in time.

The ligament size λ\lambda, extracted from the structure factor of the α\alpha–liquid interface (Supplementary Note 3), follows the Lifshitz–Slyozov–Wagner cubic coarsening law[32, 33]

λ3−λ(0)3=k(t−t0),(1)\lambda^{3} - \lambda(0)^{3} = k\left(t-t_{0}\right), \tag*{(1)}

with a fitted coarsening constant k=4.7 μm3/sk = 4.7\ \mathrm{\mu m}^{3}/\mathrm{s} at superheating ΔT=100 K\Delta T = 100\ \mathrm{K} (Figure 3c). This value predicts a ligament size of approximately 6.5 μm6.5\ \mathrm{\mu m} after one minute of dealloying, consistent with experimental observations of Ti–Ag PM[26, 27]. Sub-volumes closer to the initial α\alpha nucleus have larger λ\lambda at a given time because those ligaments entered the coarsening regime earlier (inset of Figure 3c), providing an internal age gradient within the same simulation.

Combining topology and length scale through the dimensionless quantity (g−1)λ3/V(g-1)\lambda^{3}/V, where VV is the analyzed volume, yields a value approaching ∼2\sim2 at long times (Figure 3d). This near-constant scaled genus indicates self-similar coarsening of the bicontinuous structure and is consistent with previous PF studies of nanoporous-metal coarsening[46]. The same behavior is obtained at ΔT=40 K\Delta T = 40\ \mathrm{K} (Supplementary Figure 2), where g≫1g \gg1, λ3\lambda^{3} grows linearly in time, and the scaled genus approaches a similar value, showing that bicontinuity and self-similar coarsening are robust consequences of the LFM instability.

Supplementary Figure 2. Topology and coarsening of the bicontinuous α\alpha/liquid network at ΔT=40\Delta T = 40 K. Three-dimensional PF simulation initiated from a Ti-rich α\alpha nucleus immersed in the Ag-rich liquid film between neighboring parent β\beta grains; curve colors in (b–d) correspond to the analysis sub-volumes in Figure 3a. (a) Hyperbranched α\alpha network, with the α\alpha–liquid interface shown in brown and the liquid and β\beta phases rendered transparent. (b) Evolution of g−1g - 1, where gg is the genus of the α\alpha–liquid interface in each sub-volume. Values g−1≫1g - 1 \gg1 confirm that the dealloyed region is topologically bicontinuous, and the subsequent decrease reflects capillarity-driven coarsening; the inset shows the final α\alpha morphology. (c) Cube of ligament size, λ3\lambda^3, versus time. The curves are shifted in time so as to align λ\lambda across sub-volumes. The dashed line is a linear fit, giving k≃4.7 μm3/sk \simeq4.7\ \mu\mathrm{m}^3/\mathrm{s}, consistent with the value at ΔT=100\Delta T = 100 K (Figure 3c). (d) Scaled genus, (g−1)λ3/V(g - 1)\lambda^3/V, with VV the sub-volume, saturating near ∼2\sim2 at long times and indicating robust self-similar coarsening.

Tip kinetics

A central distinction between PM and conventional LMD lies in the transport geometry. In LMD, the dissolved component must diffuse through the growing dealloyed layer to reach an external liquid medium, which leads to time-dependent kinetics, typically v∝1/tv \propto1/\sqrt{t}[15, 20]. In PM, no external reservoir exists. The Ag-rich liquid is generated locally at the melting front, and Ag rejected by the growing α\alpha phase is consumed locally by the receding β\beta–liquid interface. This confines solute redistribution to the migrating liquid film. To investigate the resulting growth kinetics, we performed moving-frame simulations in which the computational domain follows the dealloying front (Figure 4a,Figure 4b). This geometry allows the front velocity and length scales to be measured over long times. The multiple-dendrite LFM front contains many interacting α\alpha fingers, but the averaged quantities are well defined: the liquid-film thickness h′h', tip radius R′R', front-plane ligament size λ0\lambda_0, and growth velocity vv all fluctuate around steady mean values. Their approximate time independence is the kinetic signature of local solute redistribution at the PM front and contrasts sharply with the diffusion-limited, time-dependent kinetics of LMD.

Figure 4. Tip kinetics during peritectic melting. (a) Moving-frame simulation of the multiple-dendrite LFM front. (b) The liquid-film thickness h′h', tip radius R′R', front-plane ligament size λ0\lambda_0, and velocity vv remain nearly time-independent. (c) Single-dendrite LFM simulation. (d) In the single-dendrite case, h′h' and R′R' are measured directly from the interface, while hh and RR are obtained from a parabolic fit.

Although the multiple-dendrite front forms the bicontinuous network, it is too noisy for direct comparison with a tip-selection theory because individual fingers compete, branch, and coalesce. We therefore isolated the elementary growth unit by simulating a single anisotropy-selected α\alpha dendrite advancing through the same liquid-film geometry (Figure 4c,Figure 4d and Supplementary Figure 3). This reduced configuration assumes that single-dendrite growth kinetics is representative of the average behavior of the multiple-dendrite array, providing a steady LFM tip with well-defined velocity, liquid-film thickness, and curvature that can be directly compared with sharp-interface theory. The fact that isolated dendrites and multiple-dendrite arrays exhibit the same velocity-superheating relationship supports this assumption.

Parabolic fitting of the $\alpha$–liquid tip

Supplementary Figure 3. Parabolic fitting of the α\alpha–liquid tip. Analysis of a steady single α\alpha dendrite in three dimensions at ΔT=80\Delta T = 80 K. (a) Contours of the α\alpha–liquid interface at distances zz behind the tip; rr is the in-plane radial distance from the contour center. (b) Cross-sectional area SS of the α\alpha phase versus zz. For a paraboloidal tip, S=2πRzS = 2\pi Rz, so the slope gives the tip radius of curvature RR. (c) Mean cross-sectional radius rˉ=(S/π)1/2\bar{r} = (S/\pi)^{1/2} versus zz, fitted with the parabolic profile using the value of RR from (b). The fit determines the parabola-tip position and hence the liquid-film thickness hh, measured from this fitted tip to the local β\beta–liquid interface. Fits using only the three points nearest the tip and using all available data give appreciably different values of RR and tip position, illustrating the sensitivity of the extracted tip-scale quantities to the fitting range. The fit using all available data is used because it better represents the assumed paraboloidal α\alpha–liquid interface.

The superheating dependence of these kinetic quantities is summarized in Figure 5. For both single- and multiple-dendrite simulations, the growth velocity follows v∝ΔT2v \propto\Delta T^2, while the liquid-film thickness, tip radius, and initial ligament size scale approximately as ΔT−1\Delta T^{-1}. The agreement of these scalings across single and multiple (i.e. array of) primary dendrite branches indicates that the bicontinuous front inherits the same tip-scale selection physics branch by branch, even though branch interactions determine the final network topology. To model the scalings obtained in our PF simulations, we extend the existing sharp-interface theory of dendritic growth[30, 34] to LFM in 3D; previous studies only considered this extension in 2D[29, 35, 36] (Supplementary Figure 4). The theory combines two key ingredients: a solution of the diffusion equation in a frame translating at velocity vv with mass conservation conditions at the moving α\alpha– and β\beta–liquid interfaces assumed to have steady-state parabolic shapes in this frame, which is a direct extension of Ivantsov’s transport theory for a single interface[34], and a solvability condition that takes into account the effect of anisotropy[30]. The diffusion equation solution yields the relations

Scaling laws and solvability-type selection

Figure 5. Scaling laws and solvability-type selection. Symbols denote PF measurements for single and multiple dendrites; lines show fits or theoretical predictions. (a) v∝ΔT2v \propto\Delta T^2. (b–d) Film thickness, tip radius, and front-plane ligament size scale as ΔT−1\Delta T^{-1}. (e) hv/Dlhv/D_l and Pe=Rv/(2Dl)\mathrm{Pe}=Rv/(2D_l) agree with Eq. (2). (f) The stability parameter σ=d0/(PeR)\sigma=d_0/(\mathrm{Pe}R) is approximately constant.

Supplementary Figure 4. Steady-state tip kinetics in two- and three-dimensional PF simulations. Blue squares and orange circles denote 2D and 3D simulations, respectively, of a single α\alpha dendrite growing through the Ag-rich liquid film separating it from the parent β\beta grain. Here, DlD_l is the liquid diffusivity and d0d_0 is the capillary length. (a) Growth velocity vv versus superheating ΔT\Delta T, showing v∝ΔT2v \propto\Delta T^2 in both dimensions; the inset shows a steady 3D dendrite at ΔT=80\Delta T = 80 K. (b) Liquid-film thickness hh, measured from the fitted parabolic tip to the local β\beta–liquid interface, versus ΔT\Delta T; the inset defines hh and the directly measured distance h′h' in 2D. (c) Tip radius RR obtained from the parabolic fit versus ΔT\Delta T. Both hh and RR scale as ΔT−1\Delta T^{-1}. (d) Dimensionless velocity hv/Dlhv/D_l versus ΔT\Delta T, compared with the analytical predictions from Supplementary Note 4: hv/Dl=(1+ζ)δ/2hv/D_l = (1 + \zeta)\delta/2 in 2D and hv/Dl=(ζ−1)δ/ln⁡ζhv/D_l = (\zeta- 1)\delta/\ln\zeta in 3D, where δ\delta is the driving force and ζ\zeta is a phase-diagram constant. (e) Peclet number Pe=Rv/(2Dl)\mathrm{Pe} = Rv/(2D_l) versus ΔT\Delta T, compared with Pe=δ/[2(ζ−1)]\mathrm{Pe} = \delta/[2(\zeta- 1)] in 2D and Pe=δ/ln⁡ζ\mathrm{Pe} = \delta/\ln\zeta in 3D. (f) Microsolvability stability parameter σ=d0/(Pe R)\sigma= d_0/(\mathrm{Pe}\,R) versus ΔT\Delta T. The nearly constant values, σ≈0.14\sigma\approx0.14 in 2D and σ≈0.04\sigma\approx0.04 in 3D, are consistent with solvability theory [30], where σ\sigma is controlled by the interfacial anisotropy; here ϵ1=0.1\epsilon_1 = 0.1.

hvDl=(ζ−1)δln⁡ζ,Pe=δln⁡ζ,(2)\frac{hv}{D_l}=\frac{(\zeta-1)\delta}{\ln\zeta},\qquad\mathrm{Pe}=\frac{\delta}{\ln\zeta}, \tag*{(2)}

where Pe=vR/(2Dl)\mathrm{Pe}=vR/(2D_l) is the Peclet number, which is assumed small, Pe≪1\mathrm{Pe}\ll1, δ=(clα−clβ)/(cp−cpα)\delta=(c_l^\alpha-c_l^\beta)/(c_p-c_p^\alpha) is the dimensionless driving force fixed by the superheating ΔT\Delta T, and ζ=(cp−cpα)/(cp−cpβ)\zeta=(c_p-c_p^\alpha)/(c_p-c_p^\beta) is fixed by the Ti–Ag phase diagram. Since δ∼ΔT\delta\sim\Delta T near TpT_p, Eq. (2) predicts hv/Dl∝ΔThv/D_l\propto\Delta T and Pe∝ΔT\mathrm{Pe}\propto\Delta T, in agreement with the single-dendrite simulations (Figure 5e). The solvability condition, in turn, predicts that σ=d0/(PeR)\sigma=d_0/(\mathrm{Pe}R) is a constant determined by the magnitude ϵ1\epsilon_1 of anisotropy, consistent with the nearly constant value of σ≈0.04\sigma\approx0.04 extracted from 3D simulations (Figure 5f). These results yield

R∝d0σδ,v∝Dlσδ2d0,(3)R\propto\frac{d_0}{\sigma\delta},\qquad v\propto\frac{D_l\sigma\delta^2}{d_0}, \tag*{(3)}

thereby explaining the observed R∼h∼ΔT−1R\sim h\sim\Delta T^{-1} and v∼ΔT2v\sim\Delta T^2 scalings. Full derivations and definitions of d0d_0, δ\delta, and ζ\zeta are given in Supplementary Note 4.

Mechanistic implications and outlook

These results establish PM as a self-contained dealloying pathway that is mechanistically distinct from ECD, VPD, and especially LMD. Like LMD, PM involves two solid phases and a liquid metal and produces a high-genus solid/liquid morphology that coarsens self-similarly. Unlike LMD, however, PM requires no external dealloying bath: the liquid is generated internally by decomposition of the peritectic alloy, and Ag redistribution is confined to the migrating liquid film. This local transport explains why the PM front velocity and characteristic length scales remain approximately constant in time, whereas LMD exhibits time-dependent kinetics controlled by long-range diffusion through the dealloyed layer.

PM also provides a natural route to polycrystalline bicontinuous structures. Nucleation at parent β\beta grain boundaries controls the number, spacing, and orientation of α\alpha grains, whereas morphological instability of LFM, branch coalescence, and coarsening determine the ligament scale within each grain. This decoupling is absent in many LMD systems (e.g., TaTi dealloyed by a Cu melt[15]) where the secondary phase inherits the crystallographic orientation of the precursor alloy (primary phase), and bicontinuous structure formation is mediated by diffusion-limited coupled growth of solid and liquid phases rather than LFM[15]. In contrast, in LMD systems where the crystal structures of the primary and secondary phases differ, polycrystalline bicontinuous networks typically form as in (FeCr)70_{70}Ni30_{30} dealloyed in liquid Mg[47] and in SiC dealloyed in liquid Ge[19]. In PM, where the parent and product solids are always distinct phases, grain size and ligament size are determined by independent mechanisms, whereas how these scales are selected in the analogous LMD systems remains an open question.

Beyond Ti–Ag, unstable LFM together with interface coalescence may underlie bicontinuous structure formation in other systems with a dealloying medium. For example, the formation of a bicontinuous structure of graphite and a Si-rich phase observed in SiC dealloyed by liquid Ge has been interpreted within the LMD framework[19], but the mechanism has not been explicitly modeled to date. The present results suggest that the formation mechanism may be similar to the one elucidated here since the primary role of Ge is to lower the peritectic transformation temperature of SiC–Ge into graphite and a binary liquid; isothermal melting of SiC (without Ge) would be expected to form bicontinuous structures via LFM by a similar mechanism as Ti–Ag, albeit at much higher temperature. In this scenario, the presence of a dealloying medium (liquid Ge) would alter the dealloying kinetics with v∼1/tv \sim1/\sqrt{t} instead of a constant velocity, but topology would be generated by unstable LFM together with interface coalescence instead of unstable two-phase coupled growth[15]. The grain size, in turn, would be controlled by nucleation potency independently of the ligament size. The present results enlarge the scope of possible self-organized mechanisms of bicontinuous structure formation during dealloying. Which mechanism is at play in different multicomponent alloy systems remains to be investigated.

Conclusions

Phase-field simulations demonstrate that isothermal melting of an initially homogeneous Ti–Ag peritectic alloy produces a high-genus bicontinuous structure through two sequential mechanisms: a morphological instability of liquid film migration that generates a highly branched structure of the Ti-rich secondary phase, followed by interface coalescence that transforms its topology. The simulations further reveal that the solidification front velocity and the initial ligament width at this front are constant in time, with subsequent coarsening following a t1/3t^{1/3} law that predicts a final ligament width in quantitative agreement with experiment. The front velocity and initial ligament width follow scaling laws as functions of superheating in quantitative agreement with sharp-interface theory. These results establish PM as a distinct, internally driven dealloying pathway to bicontinuous structures, with grain size and ligament scale governed by independent mechanisms that can be controlled for microstructure design. More broadly, unstable LFM may underlie bicontinuous structure formation in other alloy systems even in the presence of a dealloying medium and decreasing velocity.

Acknowledgments

This work was funded by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences under Award Numbers DE-SC0020895. A.K. wishes to thank Zhongyang Li for valuable discussions. All numerical simulations were carried out on the Explorer cluster of Northeastern University at Massachusetts Green High Performance Computing Center (MGHPCC) in Holyoke, MA.

Declaration of Competing Interests

The authors report there are no competing interests to declare.

Use of AI tools

The authors used Claude Opus 4.7 (Anthropic) in two ways during preparation of this manuscript. First, for copyediting and language refinement of the main text and supplementary notes. Second, for coding assistance with the post-processing scripts used in the topology and structure-factor analyses. The tool was not used to generate scientific content, results, figures, data, or references. After using this tool, the authors reviewed and edited all output and take full responsibility for the integrity, accuracy, and originality of the content of this publication.

Supplemental Materials: Dealloying by peritectic melting

Mingwang Zhong; Alain Karma

Supplementary Note 1: Phase-field modeling of peritectic melting

We use the multi-phase-field model of Folch and Plapp [37] adapted to the peritectic decomposition of Ti50_{50}Ag50_{50}, in which the intermetallic β\beta phase melts above TpT_p into a Ti-rich solid α\alpha and an Ag-rich liquid. The state of the system is described by three non-conserved order parameters ϕα\phi_\alpha, ϕβ\phi_\beta, and ϕl\phi_l representing the local volume fractions of the three phases under the constraint ϕα+ϕβ+ϕl=1\phi_\alpha+\phi_\beta+\phi_l=1, together with the conserved Ag concentration cc. We adopt the index convention {α,β,l}→{1,2,3}\{\alpha,\beta,l\}\to\{1,2,3\} throughout. The free-energy functional is

F=∫V[∑iσ2∣∇ϕi∣2+Hfϕ(ϕ⃗)+Xfc(c,ϕ⃗)] dV,(4)\mathcal{F}=\int_V\left[\sum_i\frac{\sigma}{2}|\nabla\phi_i|^2+Hf_\phi(\vec{\phi})+Xf_c(c,\vec{\phi})\right]\,\mathrm{d}V, \tag*{(4)}

where σ\sigma is the gradient-energy coefficient that fixes the diffuse-interface width W=σ/HW=\sqrt{\sigma/H}, and HH and XX are energy-density scales. The multiwell potential

fϕ=∑iϕi2(1−ϕi)2+∑iaiϕj2ϕk2(2ϕjϕk+3ϕi+bϕi2)(5)f_\phi=\sum_i\phi_i^2(1-\phi_i)^2+\sum_i a_i\phi_j^2\phi_k^2(2\phi_j\phi_k+3\phi_i+b\phi_i^2) \tag*{(5)}

has minima at the three bulk phases. The coefficient aia_i controls the interfacial free energy of the jkjk interface that does not involve phase ii, and bb raises the energy barrier at the triple junction, suppressing the spurious emergence of the third phase at any two-phase interface. The chemical free-energy density couples ϕ⃗\vec{\phi} to cc through

fc=12[c−∑iA^i(T)gi(ϕ⃗)]2+∑iB^i(T)gi(ϕ⃗),(6)f_c=\frac{1}{2}\left[c-\sum_i\hat{A}_i(T)g_i(\vec{\phi})\right]^2+\sum_i\hat{B}_i(T)g_i(\vec{\phi}), \tag*{(6)}

with

gi(ϕ⃗)=ϕi4{15(1−ϕi)[1+ϕi−(ϕj−ϕk)2]+ϕi(9ϕi2−5)}(7)g_i(\vec{\phi})=\frac{\phi_i}{4}\left\{15(1-\phi_i)\left[1+\phi_i-(\phi_j-\phi_k)^2\right]+\phi_i(9\phi_i^2-5)\right\} \tag*{(7)}

the lowest-order interpolation function that is antisymmetric about ϕi=1/2\phi_i=1/2 and satisfies gi(ϕi=0)=0g_i(\phi_i=0)=0, gi(ϕi=1)=1g_i(\phi_i=1)=1. The temperature-dependent coefficients A^i(T)\hat{A}_i(T) and B^i(T)\hat{B}_i(T) are chosen so that Eq. (6) reproduces the linearized Ti–Ag phase diagram near TpT_p:

A^l=A1,B^l=B1(T),A^α=−A1−A2,B^α=−B1(T)−B2(T),A^β=−A1+A2,B^β=−B1(T)+B2(T),(8)\begin{aligned} \hat{A}_l&=A_1, & \hat{B}_l&=B_1(T),\\ \hat{A}_\alpha&=-A_1-A_2, & \hat{B}_\alpha&=-B_1(T)-B_2(T),\\ \hat{A}_\beta&=-A_1+A_2, & \hat{B}_\beta&=-B_1(T)+B_2(T), \tag*{(8)} \end{aligned}

with

A1=14(1+r),A2=12(1−r),B1=B11+B12T~,B2=B21+B22T~.(9)\begin{aligned} A_1&=\frac{1}{4}(1+r), & A_2&=\frac{1}{2}(1-r),\\ B_1&=B_{11}+B_{12}\widetilde{T}, & B_2&=B_{21}+B_{22}\widetilde{T}. \tag*{(9)} \end{aligned}

Here r=c~p=(c−cpβ)/(cp−cpα)∣c=cpr=\widetilde{c}_p=(c-c_{p\beta})/(c_p-c_{p\alpha})|_{c=c_p} is the scaled liquidus concentration at TpT_p and T~=(T−Tp)/[∣mα∣(cp−cpα)]\widetilde{T}=(T-T_p)/[|m_\alpha|(c_p-c_{p\alpha})] is the scaled temperature. The phase-diagram parameters TpT_p, cpαc_{p\alpha}, cpβc_{p\beta}, cpc_p, mαm_\alpha, and mβm_\beta are listed in Supplementary Table 1 and visualized in Supplementary Figure S1.

Figure S1. Phase diagram of the Ti-Ag system around the peritectic temperature TpT_p. Parameters are given in Supplementary Table 1.

The conserved field evolves by a Cahn–Hilliard equation,

∂c∂t=∇⋅[M(ϕ⃗)∇μ],M(ϕ⃗)=DlϕlH,μ=δFδc,(10)\frac{\partial c}{\partial t}=\nabla\cdot[M(\vec{\phi})\nabla\mu],\qquad M(\vec{\phi})=\frac{D_l\phi_l}{H},\qquad\mu=\frac{\delta\mathcal{F}}{\delta c}, \tag*{(10)}

in which the atomic mobility vanishes linearly with ϕl\phi_l at the solid–liquid interfaces. The non-conserved phase fields follow a constrained Allen–Cahn equation,

τ(ϕ⃗)∂ϕi∂t=−1H(δFδϕi−13∑jδFδϕj),(11)\tau(\vec{\phi})\frac{\partial\phi_i}{\partial t}=-\frac{1}{H}\left(\frac{\delta\mathcal{F}}{\delta\phi_i}-\frac{1}{3}\sum_j\frac{\delta\mathcal{F}}{\delta\phi_j}\right), \tag*{(11)}

in which the second term enforces ∑iϕi=1\sum_i \phi_i = 1 and the position-dependent time scale τ(ϕ⃗)\tau(\vec{\phi}) is chosen to eliminate the residual interface-thickness dependence of the simulated kinetics. Setting the effective interface kinetic coefficient at the α\alpha–liquid and β\beta–liquid interfaces to zero in the thin-interface limit (Eq. (3.41) of Folch and Plapp[37]) fixes two interface-specific time scales,

τα=0.7464λ~∣A^l−A^α∣2W2Dl,τβ=0.7464λ~∣A^l−A^β∣2W2Dl,(12)\tau_\alpha= 0.7464\widetilde{\lambda}|\widehat{A}_l-\widehat{A}_\alpha|^2\frac{W^2}{D_l}, \qquad\tau_\beta= 0.7464\widetilde{\lambda}|\widehat{A}_l-\widehat{A}_\beta|^2\frac{W^2}{D_l}, \tag*{(12)}

where the numerical prefactor 0.7464 comes from the matched asymptotic analysis of Folch and Plapp[37], and we define the average τ‾=(τα+τβ)/2\overline{\tau}=(\tau_\alpha+\tau_\beta)/2. To ensure that τ(ϕ⃗)\tau(\vec{\phi}) reduces to τα\tau_\alpha inside the α\alpha bulk, τβ\tau_\beta inside the β\beta bulk, and τ‾\overline{\tau} inside the liquid, we follow Eq. (3.46) of Folch and Plapp[37]:

τ(ϕ⃗)={τ‾+(τβ−τα)(ϕβ−ϕα)2(ϕα+ϕβ),if ϕl≠1,τ‾,if ϕl=1.(13)\tau(\vec{\phi}) = \begin{cases} \overline{\tau}+\dfrac{(\tau_\beta-\tau_\alpha)(\phi_\beta-\phi_\alpha)}{2(\phi_\alpha+\phi_\beta)}, & \text{if } \phi_l \ne1,\\ \overline{\tau}, & \text{if } \phi_l = 1. \end{cases} \tag*{(13)}

Equations (S7) and (S8) are non-dimensionalized using WW as the length scale and τ‾\overline{\tau} as the time scale, giving x~=x/W\widetilde{x}=x/W, t~=t/τ‾\widetilde{t}=t/\overline{\tau}, ∇~=W∇\widetilde{\nabla}=W\nabla, μ~=μ/H\widetilde{\mu}=\mu/H, F~=F/H\widetilde{F}=F/H, and the dimensionless chemical prefactor λ~=X/H\widetilde{\lambda}=X/H. The dimensionless equations are integrated on a uniform Cartesian grid with second-order central finite differences and explicit Euler time stepping. The time step is bounded by the diffusion stability limit,

dt~=0.6 dx~22ddimD~l,D~l=Dlτ‾W2,(14)\widetilde{dt}=\frac{0.6\,d\widetilde{x}^{2}}{2d_{\mathrm{dim}}\widetilde{D}_l}, \qquad\widetilde{D}_l=\frac{D_l\overline{\tau}}{W^2}, \tag*{(14)}

where ddim=2d_{\mathrm{dim}}=2 in 2D and ddim=3d_{\mathrm{dim}}=3 in 3D arises from the central-difference Laplacian of the explicit scheme, and the prefactor 0.6 is chosen small enough to satisfy stability and large enough to accelerate the simulations. The code is implemented in CUDA C/C++; simulations are carried out on NVIDIA V100 or H200 GPUs.

ParameterDescriptionValue
TpT_pPeritectic temperature of the Ti-Ag system1293 K
cpαc_{p\alpha}Ag composition of the α\alpha solidus at peritectic temperature16.41 at.%
cpβc_{p\beta}Ag composition of the β\beta solid at peritectic temperature50 at.%
cpc_pAg composition of the liquidus at peritectic temperature93.95 at.%
mαm_\alphaLiquidus slope above peritectic temperature-40.078 K/at.%
mβm_\betaLiquidus slope below peritectic temperature-26.187 K/at.%
DlD_lDiffusion constant in the liquid2×10−92 \times10^{-9} m2^2/s
γαl\gamma_{\alpha l}Interfacial free-energy for the liquid-α\alpha interface0.20 J/m2^2
γβl\gamma_{\beta l}Interfacial free-energy for the liquid-β\beta interface0.19 J/m2^2
γαβ\gamma_{\alpha\beta}Interfacial free-energy for the α\alpha-β\beta interface0.18 J/m2^2

Table 1. Parameters of the phase-field model

Estimation of model parameters

Interfacial free energies, aia_i and bb. For each interface, the dimensionless surface free energy is set by the multiwell coefficient associated with the third phase. With {α,β,l}→{1,2,3}\{\alpha,\beta,l\}\to\{1,2,3\}, γβl\gamma_{\beta l} is controlled by a1a_1, γαl\gamma_{\alpha l} by a2a_2, and γαβ\gamma_{\alpha\beta} by a3a_3, through[37]

\widetilde{\gamma}_{jk}(a_i)=\frac{2\sqrt{a_i}(3a_i+4)+(a_i+4)(3a_i-4)\arccot(2/\sqrt{a_i})}{16\sqrt{2}\,a_i^{3/2}}, \qquad\gamma_{jk}=\widetilde{\gamma}_{jk}WH. \tag*{(15)}

We use a single set of interfacial free energies throughout, γβl=0.19\gamma_{\beta l}=0.19, γαl=0.20\gamma_{\alpha l}=0.20, and γαβ=0.18 J/m2\gamma_{\alpha\beta}=0.18\ \mathrm{J}/\mathrm{m}^{2}, corresponding to a1=4.54285a_1=4.54285, a2=5.58545a_2=5.58545, and a3=3.55521a_3=3.55521, respectively. The triple-junction coefficient is fixed at b=80b=80, large enough to fully suppress the spurious appearance of any third phase at a two-phase interface.

Energy-density scale and λ~\tilde{\lambda}. The dimensionless prefactor λ~=X/H\tilde{\lambda}=X/H is introduced as the relative weight of chemical and gradient-energy contributions in Eq. (4). Equivalently, in the scaled variables c~=(c−cpβ)/(cp−cpα)\tilde{c}=(c-c_{p\beta})/(c_p-c_{p\alpha}) and μ~=μ/H\tilde{\mu}=\mu/H, it is the slope of the equilibrium chemical potential with respect to concentration,

λ~=∂μ~∂c~∣eq(16)\tilde{\lambda}=\left.\frac{\partial\tilde{\mu}}{\partial\tilde{c}}\right|_{\mathrm{eq}} \tag*{(16)}

We calibrate λ~\tilde{\lambda} by matching the latent heat of melting across the α\alpha–liquid and β\beta–liquid interfaces at TpT_p, which we compute in two independent ways and equate.

Route 1 — thermodynamic functions. Equilibrium of phase ii and liquid at TpT_p requires equal molar free energies,

Glm−Gim=0=Hlm−Him−Tp(Slm−Sim)=Lim−TpΔSim,(17)G_l^m-G_i^m=0=H_l^m-H_i^m-T_p\left(S_l^m-S_i^m\right)=L_i^m-T_p\Delta S_i^m, \tag*{(17)}

where Lim=Hlm−HimL_i^m=H_l^m-H_i^m is the latent heat of melting of phase ii. Using the Gibbs–Helmholtz identity Sm=−∂Gm/∂TS^m=-\partial G^m/\partial T,

Lim=Tp(Slm−Sim)=Tp(∂Glm∂T∣p−∂Gim∂T∣p),(18)L_i^m=T_p\left(S_l^m-S_i^m\right)=T_p\left(\left.\frac{\partial G_l^m}{\partial T}\right|_p-\left.\frac{\partial G_i^m}{\partial T}\right|_p\right), \tag*{(18)}

which, evaluated with the published temperature-dependent free-energy functions Glm(T)G_l^m(T), Gαm(T)G_\alpha^m(T), Gβm(T)G_\beta^m(T) of the Ti–Ag binary [39, 40], yields Lα=15.9 kJ/molL_\alpha=15.9\ \mathrm{kJ/mol} and Lβ=28.3 kJ/molL_\beta=28.3\ \mathrm{kJ/mol}.

Route 2 — phase-field model. The same latent heat can be obtained from the grand-potential balance across the ii–liquid interface implied by Eqs. (4)–(6). Defining the molar volume VmV_m and the equilibrium concentration jump Δcα=cp−cpα\Delta c_\alpha=c_p-c_{p\alpha}, the molar free-energy difference takes the compact form

Glm−Gim=μ~liΔcα(c~lli−c~ili)HVm,(19)G_l^m-G_i^m=\tilde{\mu}^{li}\Delta c_\alpha\left(\tilde{c}_l^{li}-\tilde{c}_i^{li}\right)H V_m, \tag*{(19)}

where c~lli\tilde{c}_l^{li} and c~ili\tilde{c}_i^{li} are the scaled equilibrium concentrations of the liquid and the ii solid across the interface, and μ~li\tilde{\mu}^{li} is the scaled equilibrium chemical potential at that interface (we omit VmV_m in what follows since it cancels with the molar normalization on the left-hand side). Differentiating Eq. (19) along the equilibrium phase boundary at TpT_p and using Eq. (18) gives

Li=TpΔcα∂μ~∂T∣p,eq(c~lli−c~ili)H=TpΔcα∂μ~∂c~∣p,eq∂c∂T∣p,eq(c~lli−c~ili)H,\begin{aligned} L_i&=T_p\Delta c_\alpha\left.\frac{\partial\tilde{\mu}}{\partial T}\right|_{p,\mathrm{eq}}\left(\tilde{c}_l^{li}-\tilde{c}_i^{li}\right)H\\ &=T_p\Delta c_\alpha\left.\frac{\partial\tilde{\mu}}{\partial\tilde{c}}\right|_{p,\mathrm{eq}}\left.\frac{\partial c}{\partial T}\right|_{p,\mathrm{eq}}\left(\tilde{c}_l^{li}-\tilde{c}_i^{li}\right)H, \end{aligned}

where the second line applies the chain rule. The two factors on the second line are evaluated as follows. Along the linearized liquidus on the ii side of TpT_p, T=Tp+mi(c−cp)T=T_p+m_i(c-c_p) with slope mi<0m_i<0, so

∂c∂T∣p,eq=1mi,i.e.,∣∂c∂T∣p,eq∣=1∣mi∣.(20)\left.\frac{\partial c}{\partial T}\right|_{p,\mathrm{eq}}=\frac{1}{m_i},\qquad\text{i.e.,}\qquad\left|\left.\frac{\partial c}{\partial T}\right|_{p,\mathrm{eq}}\right|=\frac{1}{|m_i|}. \tag*{(20)}

For the chemical-potential slope, the definition c~=(c−cpβ)/Δcα\tilde{c}=(c-c_{p\beta})/\Delta c_\alpha gives ∂c~/∂c=1/Δcα\partial\tilde{c}/\partial c=1/\Delta c_\alpha, so Eq. (16) implies

∂μ~∂c∣p,eq=1Δcα∂μ~∂c~∣eq=λ~Δcα.(21)\left.\frac{\partial\tilde{\mu}}{\partial c}\right|_{p,\mathrm{eq}}=\frac{1}{\Delta c_\alpha}\left.\frac{\partial\tilde{\mu}}{\partial\tilde{c}}\right|_{\mathrm{eq}}=\frac{\tilde{\lambda}}{\Delta c_\alpha}. \tag*{(21)}

Substituting Eqs. (20) and (21) into Eq. (20) and using c~lli>c~ili\tilde{c}_{l}^{li} > \tilde{c}_{i}^{li} to keep LiL_i positive, the Δcα\Delta c_{\alpha} factors cancel and we arrive at

Li=Tpλ~∣mi∣(c~lli−c~ili)H.(22)L_i = \frac{T_p\tilde{\lambda}}{\lvert m_i\rvert}\left(\tilde{c}_{l}^{li} - \tilde{c}_{i}^{li}\right)H. \tag*{(22)}

Equation (S19) provides one equation per interface relating λ~\tilde{\lambda} to the latent heat. Using Lα=15.9L_{\alpha} = 15.9 kJ/mol and Lβ=28.3L_{\beta} = 28.3 kJ/mol from Route 1, the parameters in Supplementary Table 1, and the equilibrium concentrations c~lα\tilde{c}_{l}^{\alpha}, c~αα\tilde{c}_{\alpha}^{\alpha}, c~lβ\tilde{c}_{l}^{\beta}, c~ββ\tilde{c}_{\beta}^{\beta} read directly from the Folch–Plapp phase diagram at TpT_p, Eq. (22) yields λ~≈25\tilde{\lambda} \approx25 from the α\alpha–liquid interface and λ~≈51\tilde{\lambda} \approx51 from the β\beta–liquid interface. As a compromise that resolves both interfaces with a single prefactor and keeps the dimensionless equations well conditioned, we adopt λ~ref=30\tilde{\lambda}_{\mathrm{ref}} = 30 together with Href=1.462×108H_{\mathrm{ref}} = 1.462 \times10^{8} J/m3^{3} and Wref=2W_{\mathrm{ref}} = 2 nm as a reference parameter set; the product WH=0.2924WH = 0.2924 J/m2^{2} then reproduces the dimensional surface energies quoted above.

Width adaptation across superheatings. Because all dimensional surface energies satisfy γjk=γ~jkWH\gamma_{jk} = \tilde{\gamma}_{jk}WH, the product WHWH uniquely sets the surface tension: doubling WW and halving HH leaves γjk\gamma_{jk} unchanged. The latent heat Li∝λ~HL_i \propto\tilde{\lambda}H is preserved if λ~\tilde{\lambda} is rescaled inversely with HH, i.e. λ~→λ~(Href/H)\tilde{\lambda} \to\tilde{\lambda}(H_{\mathrm{ref}}/H). The simulated physics, including the front velocity vv, the liquid-film thickness hh, and the tip radius RR, is therefore invariant under this coupled rescaling. We exploit this freedom to enlarge WW at low driving forces, where the underlying physical length scales are intrinsically larger. For each superheating ΔT\Delta T we use

W(ΔT)=200 KΔT×2 nm=400ΔT nm,(23)W(\Delta T) = \frac{200\ \mathrm{K}}{\Delta T} \times2\ \mathrm{nm} = \frac{400}{\Delta T}\ \mathrm{nm}, \tag*{(23)}

which recovers W=2W = 2 nm at the reference ΔT=200\Delta T = 200 K and grows WW as ΔT\Delta T decreases. This choice maximizes the dimensionless grid spacing and time step (Eq. (14)) and so accelerates the low-ΔT\Delta T simulations while leaving hh, vv, and RR unchanged to within numerical tolerance.

Anisotropic α\alpha-liquid interfacial free energy

The dependence of the α\alpha–liquid interfacial free energy on the orientation of the local interface normal is incorporated only at the α\alpha–liquid interface; the β\beta–liquid and α\alpha–β\beta interfaces remain isotropic. The dimensionless anisotropy function takes the standard fourfold cubic-harmonic form [31, 41],

as(n)=1+ϵ1(∑i=x,y,zni′4−35),(24)a_s(\mathbf{n}) = 1 + \epsilon_1 \left( \sum_{i=x,y,z} {n_i'}^4 - \frac{3}{5} \right), \tag*{(24)}

where n\mathbf{n} is the unit normal to the α\alpha–liquid interface in the lab frame and ni′=n⋅e^i′n_i' = \mathbf{n} \cdot\hat{\mathbf{e}}_i' is its projection onto the α\alpha crystallographic axis i′∈{x′,y′,z′}i' \in\{x',y',z'\} specified by the orthonormal unit vectors x′=(x1,x2,x3)\mathbf{x}' = (x_1,x_2,x_3), y′=(y1,y2,y3)\mathbf{y}' = (y_1,y_2,y_3), and z′=(z1,z2,z3)\mathbf{z}' = (z_1,z_2,z_3). We use ϵ1=0\epsilon_1 = 0 for the isotropic simulations and ϵ1=0.1\epsilon_1 = 0.1 for the weakly anisotropic simulations of the main text.

The anisotropic surface free energy modifies the evolution equation for ϕα\phi_\alpha in two ways [42]. First, the characteristic time on the left-hand side of Eq. (11) acquires the orientation-dependent factor

τ(ϕ⃗) ∂tϕα⟶τ(ϕ⃗)as2(n) ∂tϕα.(25)\tau(\vec{\phi}) \, \partial_t \phi_\alpha\longrightarrow\tau(\vec{\phi}) a_s^2(\mathbf{n}) \, \partial_t \phi_\alpha. \tag*{(25)}

Second, the Laplacian contribution σ∇2ϕα\sigma\nabla^2\phi_\alpha arising from the gradient-energy term of the functional derivative δF/δϕα\delta\mathcal{F}/\delta\phi_\alpha is replaced by the divergence form

∇⋅[as2(n)∇ϕα]+∑i∈{x,y,z}∂i[∣∇ϕα∣2as(n)∂as(n)∂(∂iϕα)],(26)\nabla\cdot\left[a_s^2(\mathbf{n}) \nabla\phi_\alpha\right] + \sum_{i\in\{x,y,z\}} \partial_i \left[|\nabla\phi_\alpha|^2 a_s(\mathbf{n}) \frac{\partial a_s(\mathbf{n})}{\partial(\partial_i\phi_\alpha)}\right], \tag*{(26)}

which reduces to ∇2ϕα\nabla^2\phi_\alpha when ϵ1=0\epsilon_1 = 0. All other terms in the evolution equations for ϕα\phi_\alpha, ϕβ\phi_\beta, and ϕl\phi_l are unchanged.

Two-dimensional implementation. In 2D, n=(cos⁡θ,sin⁡θ,0)\mathbf{n} = (\cos\theta,\sin\theta,0) where θ\theta is the angle between n\mathbf{n} and the lab xx-axis, so that the crystal-frame projections are

nx′=x1cos⁡θ+x2sin⁡θ,ny′=y1cos⁡θ+y2sin⁡θ,nz′=z1cos⁡θ+z2sin⁡θ.(27)n_x' = x_1\cos\theta+ x_2\sin\theta,\qquad n_y' = y_1\cos\theta+ y_2\sin\theta,\qquad n_z' = z_1\cos\theta+ z_2\sin\theta. \tag*{(27)}

Following Tourret et al. [42], the anisotropic operator Eq. (26) is expanded as

∇⋅[as2∇ϕα]+∑i=x,y∂i[∣∇ϕα∣2as∂as∂(∂iϕα)]=as2∇2ϕα+2asas′(∂xϕα ∂xθ+∂yϕα ∂yθ)+(∂xϕα ∂yθ−∂yϕα ∂xθ)(asas′′+(as′)2),(28)\begin{aligned} \nabla\cdot\left[a_s^2 \nabla\phi_\alpha\right] + \sum_{i=x,y} \partial_i\left[|\nabla\phi_\alpha|^2 a_s \frac{\partial a_s}{\partial(\partial_i\phi_\alpha)}\right] ={}& a_s^2\nabla^2\phi_\alpha+ 2a_sa_s'(\partial_x\phi_\alpha\,\partial_x\theta+\partial_y\phi_\alpha\,\partial_y\theta)\\ &+(\partial_x\phi_\alpha\,\partial_y\theta-\partial_y\phi_\alpha\,\partial_x\theta)(a_sa_s''+(a_s')^2), \tag*{(28)} \end{aligned}

with as′≡∂as/∂θa_s' \equiv\partial a_s/\partial\theta and as′′≡∂2as/∂θ2a_s'' \equiv\partial^2a_s/\partial\theta^2 obtained by differentiating Eq. (24) through the chain rule and Eq. (27):

∂as∂θ=4ϵ1(nx′3∂nx′∂θ+ny′3∂ny′∂θ+nz′3∂nz′∂θ),\frac{\partial a_s}{\partial\theta}=4\epsilon_1\left({n_x'}^3\frac{\partial n_x'}{\partial\theta}+{n_y'}^3\frac{\partial n_y'}{\partial\theta}+{n_z'}^3\frac{\partial n_z'}{\partial\theta}\right),
∂2as∂θ2=4ϵ1[3nx′2(∂nx′∂θ)2+nx′3∂2nx′∂θ2+3ny′2(∂ny′∂θ)2+ny′3∂2ny′∂θ2+3nz′2(∂nz′∂θ)2+nz′3∂2nz′∂θ2],(29)\frac{\partial^2 a_s}{\partial\theta^2}=4\epsilon_1\left[3{n_x'}^2\left(\frac{\partial n_x'}{\partial\theta}\right)^2+{n_x'}^3\frac{\partial^2n_x'}{\partial\theta^2}+3{n_y'}^2\left(\frac{\partial n_y'}{\partial\theta}\right)^2+{n_y'}^3\frac{\partial^2n_y'}{\partial\theta^2}+3{n_z'}^2\left(\frac{\partial n_z'}{\partial\theta}\right)^2+{n_z'}^3\frac{\partial^2n_z'}{\partial\theta^2}\right], \tag*{(29)}

where the derivatives ∂ni′/∂θ\partial n_i'/\partial\theta and ∂2ni′/∂θ2\partial^2n_i'/\partial\theta^2 follow directly from Eq. (27). Because θ=tan⁡−1(∂yϕα/∂xϕα)\theta=\tan^{-1}(\partial_y\phi_\alpha/\partial_x\phi_\alpha), its first derivatives are

∂xθ=∂xyϕα ∂xϕα−∂yϕα ∂xxϕα∣∇ϕα∣2,∂yθ=∂yyϕα ∂xϕα−∂yϕα ∂xyϕα∣∇ϕα∣2.(30)\partial_x\theta=\frac{\partial_{xy}\phi_\alpha\,\partial_x\phi_\alpha-\partial_y\phi_\alpha\,\partial_{xx}\phi_\alpha}{|\nabla\phi_\alpha|^2},\qquad \partial_y\theta=\frac{\partial_{yy}\phi_\alpha\,\partial_x\phi_\alpha-\partial_y\phi_\alpha\,\partial_{xy}\phi_\alpha}{|\nabla\phi_\alpha|^2}. \tag*{(30)}

Three-dimensional implementation. In 3D we work directly with the components of the gradient. Defining p=∂xϕαp=\partial_x\phi_\alpha, q=∂yϕαq=\partial_y\phi_\alpha, u=∂zϕαu=\partial_z\phi_\alpha, and r=∣∇ϕα∣r=\lvert\nabla\phi_\alpha\rvert, the projections become

nx′=x1p+x2q+x3ur,ny′=y1p+y2q+y3ur,nz′=z1p+z2q+z3ur,(31)n_x'=\frac{x_1p+x_2q+x_3u}{r},\qquad n_y'=\frac{y_1p+y_2q+y_3u}{r},\qquad n_z'=\frac{z_1p+z_2q+z_3u}{r}, \tag*{(31)}

and their derivatives with respect to the gradient components are

∂pnx′=x1r−x1p2+x2pq+x3pur3,∂qnx′=x2r−x1pq+x2q2+x3qur3,∂unx′=x3r−x1pu+x2qu+x3u2r3,(32)\begin{aligned} \partial_p n_x'&=\frac{x_1}{r}-\frac{x_1p^2+x_2pq+x_3pu}{r^3},\\ \partial_q n_x'&=\frac{x_2}{r}-\frac{x_1pq+x_2q^2+x_3qu}{r^3},\\ \partial_u n_x'&=\frac{x_3}{r}-\frac{x_1pu+x_2qu+x_3u^2}{r^3}, \tag*{(32)} \end{aligned}

with the analogous relations for ny′n_y' and nz′n_z'. The derivatives of asa_s with respect to pp, qq, uu then take the form

∂as∂p=4ϵ1(nx′3∂pnx′+ny′3∂pny′+nz′3∂pnz′),(33)\frac{\partial a_s}{\partial p}=4\epsilon_1\left(n_x'^3\partial_p n_x'+n_y'^3\partial_p n_y'+n_z'^3\partial_p n_z'\right), \tag*{(33)}

and analogously for ∂as/∂q\partial a_s/\partial q and ∂as/∂u\partial a_s/\partial u. Substituting into Eq. (26) gives the two anisotropic contributions

∇⋅[as2(n)∇ϕα]+∑i∈{x,y,z}∂i[∣∇ϕα∣2as(n)∂as(n)∂(∂iϕα)]=as2∇2ϕα+2as(∂xϕα ∂xas+∂yϕα ∂yas+∂zϕα ∂zas)+∑q∈{x,y,z}∂qtq,(34)\begin{aligned} \nabla\cdot\left[a_s^2(\mathbf{n})\nabla\phi_\alpha\right]+\sum_{i\in\{x,y,z\}}\partial_i\left[\lvert\nabla\phi_\alpha\rvert^2a_s(\mathbf{n})\frac{\partial a_s(\mathbf{n})}{\partial(\partial_i\phi_\alpha)}\right] \\ &=a_s^2\nabla^2\phi_\alpha+2a_s\left(\partial_x\phi_\alpha\,\partial_xa_s+\partial_y\phi_\alpha\,\partial_ya_s+\partial_z\phi_\alpha\,\partial_za_s\right)+\sum_{q\in\{x,y,z\}}\partial_qt_q, \tag*{(34)} \end{aligned}

where

tx=∣∇ϕα∣2as∂as∂p,ty=∣∇ϕα∣2as∂as∂q,tz=∣∇ϕα∣2as∂as∂u.(35)t_x=\lvert\nabla\phi_\alpha\rvert^2a_s\frac{\partial a_s}{\partial p},\qquad t_y=\lvert\nabla\phi_\alpha\rvert^2a_s\frac{\partial a_s}{\partial q},\qquad t_z=\lvert\nabla\phi_\alpha\rvert^2a_s\frac{\partial a_s}{\partial u}. \tag*{(35)}

Numerical strategy. The 3D update is implemented as two CUDA kernels: the first computes the auxiliary fields asa_s, txt_x, tyt_y, tzt_z at every grid point; the second uses these fields to assemble the anisotropic contribution and advance ϕα\phi_\alpha. Boundary conditions are applied to all four auxiliary fields on the faces, edges, and vertices of the simulation domain. This split-kernel scheme avoids redundant computation and is roughly twice as fast as a single-kernel implementation in our tests. Because each α\alpha orientation requires its own auxiliary fields, the scheme is memory-intensive for polycrystalline simulations with many orientations; the present simulations use a single α\alpha orientation, so memory is not a binding constraint.

Supplementary Note 3: Calculation of genus and characteristic length

The bicontinuous topology of the α/L\alpha/L network is quantified by two complementary measures: the genus gg of the α\alpha–liquid interface, which counts independent handles, and the characteristic length λ\lambda extracted from the structure factor of the phase field, which measures the typical ligament size. We follow the marching-cubes-based methodology used by Geslin et al. [46], summarized below.

Genus from a triangulated interface. The α\alpha–liquid interface is extracted from the order parameter ϕ\phi as the level set ϕ=0\phi= 0 using the marching-cubes algorithm[38], which yields a closed triangulated surface. For a closed orientable surface composed of NblobN_{\mathrm{blob}} disjoint connected components with total genus g=∑igig = \sum_i g_i, the Euler characteristic of the triangulation obeys

χ≡NV−NE+NF=2Nblob−2g,(36)\chi\equiv N_V - N_E + N_F = 2N_{\mathrm{blob}} - 2g, \tag*{(36)}

where NVN_V, NEN_E, and NFN_F are the numbers of vertices, edges, and faces of the triangulation. Solving for gg gives

g=12(2Nblob+NE−NV−NF).(37)g = \frac{1}{2}\left(2N_{\mathrm{blob}} + N_E - N_V - N_F\right). \tag*{(37)}

Marching-cubes implementation. In the marching-cubes scheme, each grid cube is classified by the signs of ϕ\phi at its eight corners, yielding 28=2562^8 = 256 configurations that map, through a precomputed lookup table, to a fixed set of interior triangles and intersected (“cut”) cube edges. Summing over all cubes ii and correcting for sharing between neighboring cubes,

NF=∑iNtrii,NE=32∑iNtrii,NV=14∑iNCutEdgei,(38)N_F = \sum_i N_{\mathrm{tri}}^i,\qquad N_E = \frac{3}{2}\sum_i N_{\mathrm{tri}}^i,\qquad N_V = \frac{1}{4}\sum_i N_{\mathrm{CutEdge}}^i, \tag*{(38)}

where NtriiN_{\mathrm{tri}}^i is the number of triangles inside cube ii and NCutEdgeiN_{\mathrm{CutEdge}}^i is the number of cube edges that the interface crosses. The factor 3/23/2 in NEN_E reflects that each triangle contributes three edges, each shared between two triangles, and the factor 1/41/4 in NVN_V reflects that each surface vertex is shared by four neighboring cubes. Substituting Eq. (38) into Eq. (37) and simplifying gives

g=Nblob+14∑iNtrii−18∑iNCutEdgei.(39)g = N_{\mathrm{blob}} + \frac{1}{4}\sum_i N_{\mathrm{tri}}^i - \frac{1}{8}\sum_i N_{\mathrm{CutEdge}}^i. \tag*{(39)}

The two cube-level quantities NtriiN_{\mathrm{tri}}^i and NCutEdgeiN_{\mathrm{CutEdge}}^i are read directly from the 256-entry lookup tables, and NblobN_{\mathrm{blob}} is obtained by 26-connectivity component labeling of the α\alpha phase.

Characteristic length from the structure factor. The characteristic ligament size λ\lambda is extracted from the structure factor of the order parameter [46],

S(k)=∑r∑r′e−ik⋅r′[ϕ(r+r′)ϕ(r)−⟨ϕ⟩2]N2(⟨ϕ2⟩−⟨ϕ⟩2)(40)S(\mathbf{k}) = \frac{\sum_{\mathbf{r}}\sum_{\mathbf{r}'} e^{-i\mathbf{k}\cdot\mathbf{r}'}\left[\phi(\mathbf{r}+\mathbf{r}')\phi(\mathbf{r})-\langle\phi\rangle^2\right]}{N^2\left(\langle\phi^2\rangle-\langle\phi\rangle^2\right)} \tag*{(40)}

where N=NxNyNzN = N_xN_yN_z is the number of grid points in the analyzed sub-volume and ⟨⋅⟩\langle\cdot\rangle denotes the spatial average over that sub-volume. Spherical averaging over directions gives the isotropic spectrum S(k)S(k) with k=∣k∣k = |\mathbf{k}|, and the characteristic wavenumber is defined as the spectrum-weighted mean

kˉ=∫kS(k) dk∫S(k) dk,λ=2πkˉ.(41)\bar{k} = \frac{\int kS(k)\,dk}{\int S(k)\,dk},\qquad\lambda= \frac{2\pi}{\bar{k}}. \tag*{(41)}

The front-plane ligament size λ0\lambda_0 is computed analogously, with the averages in Eqs. (40)–(41) restricted to a two-dimensional cross section taken at the dealloying front.

Solvability theory of the coupled moving interfaces

This note derives the closed-form tip-scale relations [Eqs. (2)–(3)] used in the main text and provides the analytical curves plotted in Figure 5 and Supplementary Figure 4. The analysis extends the classical Ivantsov–microsolvability description of free dendritic growth [30, 34] to the LFM geometry of PM, in which an advancing α\alpha–liquid interface and a retreating β\beta–liquid interface migrate in concert across a thin Ag-rich liquid film. As in the classical theory, the diffusion field around an assumed paraboloidal α\alpha tip sets the product RvRv, i.e. the Peclet number Pe=Rv/(2Dl)\mathrm{Pe}=Rv/(2D_l), together with the film thickness hh, while microsolvability fixes the stability constant σ=d0/(Pe R)\sigma=d_0/(\mathrm{Pe}\,R) and so closes the selection problem for RR and vv separately. The new ingredient relative to free growth is that the second moving interface imposes a second concentration boundary condition behind the α\alpha tip.

Geometry and governing equation. The α\alpha–liquid tip is approximated as a paraboloid of curvature radius RR advancing along zz at constant velocity vv, and the β\beta–liquid interface as a confocal paraboloid trailing it. In parabolic coordinates normalized to RR and centred on the α\alpha tip, both interfaces are level surfaces of η\eta: the α\alpha–liquid interface lies at η=1\eta=1 and the β\beta–liquid interface at η=η0>1\eta=\eta_0>1. In the frame co-moving with the tip, the Ag concentration cc satisfies the steady advection–diffusion equation

Dl∇2c+v ∂zc=0.(42)D_l\nabla^2c+v\,\partial_zc=0. \tag*{(42)}

The Gibbs–Thomson condition at each interface,

cli=cp−T−Tp∣mi∣−Γκ∣mi∣,i=α,β,(43)c_{li}=c_p-\frac{T-T_p}{\lvert m_i\rvert}-\frac{\Gamma\kappa}{\lvert m_i\rvert},\qquad i=\alpha,\beta, \tag*{(43)}

defines capillary lengths di=Γ/[∣mi∣(cp−cpα)]d_i=\Gamma/[\lvert m_i\rvert(c_p-c_{p\alpha})]; for the Ti–Ag parameters in Supplementary Table 1 (mα=−4007.8 Km_\alpha=-4007.8\ \mathrm{K}, mβ=−2618.7 Km_\beta=-2618.7\ \mathrm{K}), dα≃3.2×10−11 md_\alpha\simeq3.2\times10^{-11}\ \mathrm{m} and dβ≃4.9×10−11 md_\beta\simeq4.9\times10^{-11}\ \mathrm{m}. For tip radii R∼10R\sim10–103 nm10^3\ \mathrm{nm}, the capillary correction to clic_{li} is at most ∼10−3\sim10^{-3}, so we ignore the curvature term when solving for the diffusion field and impose c∣η=1=clαc|_{\eta=1}=c_{l\alpha}, c∣η=η0=clβc|_{\eta=\eta_0}=c_{l\beta}, where clαc_{l\alpha} and clβc_{l\beta} are read directly from the linearized Ti–Ag phase diagram at the working temperature T=Tp+ΔTT=T_p+\Delta T. Mass conservation at each interface reads

vn(cp−cpi)=−Dl∂nc∣η=ηi,ηα=1,ηβ=η0,(44)v_n(c_p-c_{pi})=-D_l\left.\partial_nc\right|_{\eta=\eta_i},\qquad\eta_\alpha=1,\quad\eta_\beta=\eta_0, \tag*{(44)}

where vnv_n is the local normal velocity. Equations [(S39)]–[(S41)] form a closed problem for c(η)c(\eta), η0\eta_0, and Pe\mathrm{Pe}.

Two-dimensional solution. With x=Rξηx=R\xi\eta and z=R(η2−ξ2)/2z=R(\eta^2-\xi^2)/2, the scale factors are hξ=hη=Rξ2+η2h_\xi=h_\eta=R\sqrt{\xi^2+\eta^2}. Seeking an Ivantsov-type solution c=c(η)c=c(\eta) reduces Eq. (42) to

d2cdη2+2Pe ηdcdη=0,(45)\frac{d^2c}{d\eta^2}+2\mathrm{Pe}\,\eta\frac{dc}{d\eta}=0, \tag*{(45)}

with general solution

c(η)=A erf⁡(Pe η)+B,(46)c(\eta)=A\,\operatorname{erf}\left(\sqrt{\mathrm{Pe}}\,\eta\right)+B, \tag*{(46)}

where AA and BB are fixed by c(1)=clαc(1)=c_{l\alpha} and c(η0)=clβc(\eta_0)=c_{l\beta}. Taking the ratio of the two mass balances (Eq. (44)) at η=1\eta=1 and η=η0\eta=\eta_0 gives

ζ≡cp−cpαcp−cpβ=η0ePe(η02−1).(47)\zeta\equiv\frac{c_p-c_{p\alpha}}{c_p-c_{p\beta}}=\eta_0e^{\mathrm{Pe}(\eta_0^2-1)}. \tag*{(47)}

The PF simulations give Pe<0.03\mathrm{Pe} < 0.03 over the entire range ΔT≤200 K\Delta T \le200\ \mathrm{K} (Supplementary Figure 4), so η0≈ζ\eta_0 \approx\zeta. Substituting this back into the α\alpha-interface mass balance and expanding erf⁡(x)≈2x/π\operatorname{erf}(x) \approx2x/\sqrt{\pi} yields a linear concentration profile in η\eta[35] and the closed-form relations

Pe=δ2(ζ−1),h=12(ζ2−1)R,hvDl=12(ζ+1)δ,(48)\mathrm{Pe} = \frac{\delta}{2(\zeta- 1)}, \qquad h = \frac{1}{2}(\zeta^2 - 1)R, \qquad\frac{hv}{D_l} = \frac{1}{2}(\zeta+ 1)\delta, \tag*{(48)}

where the dimensionless driving force

δ≡clα−clβcp−cpα(49)\delta\equiv\frac{c_{l\alpha} - c_{l\beta}}{c_p - c_{p\alpha}} \tag*{(49)}

is linear in ΔT\Delta T near TpT_p and the film thickness follows from h=R(η02−1)/2h = R(\eta_0^2 - 1)/2.

Three-dimensional solution. With x=Rξηcos⁡ϕx = R\xi\eta\cos\phi, y=Rξηsin⁡ϕy = R\xi\eta\sin\phi, and z=R(η2−ξ2)/2z = R(\eta^2 - \xi^2)/2, axisymmetric solutions of Eq. (42) again reduce to a one-dimensional ordinary differential equation,

1ηddη(ηdcdη)+2Pe ηdcdη=0,(50)\frac{1}{\eta}\frac{d}{d\eta}\left(\eta\frac{dc}{d\eta}\right) + 2\mathrm{Pe}\,\eta\frac{dc}{d\eta} = 0, \tag*{(50)}

with general solution

c(η)=AE1(Pe η2)+B,E1(z)≡∫z∞e−tt dt.(51)c(\eta) = A E_1(\mathrm{Pe}\,\eta^2) + B, \qquad E_1(z) \equiv\int_z^\infty\frac{e^{-t}}{t}\,dt. \tag*{(51)}

Imposing the same two interfacial conditions, taking the ratio of mass balances, and using the small-argument expansion E1(z)≈γ−ln⁡z+z+O(z2)E_1(z) \approx\gamma- \ln z + z + O(z^2) give

ζ=η02ePe(η02−1)≈η02,δ≈Peln⁡ζ,(52)\zeta= \eta_0^2 e^{\mathrm{Pe}(\eta_0^2-1)} \approx\eta_0^2, \qquad\delta\approx\mathrm{Pe}\ln\zeta, \tag*{(52)}

and hence

Pe=δln⁡ζ,h=12(ζ−1)R,hvDl=(ζ−1)δln⁡ζ.(53)\mathrm{Pe} = \frac{\delta}{\ln\zeta}, \qquad h = \frac{1}{2}(\zeta- 1)R, \qquad\frac{hv}{D_l} = \frac{(\zeta- 1)\delta}{\ln\zeta}. \tag*{(53)}

Microsolvability and tip-scale selection. Equations (S45) and (S50) fix the products hvhv and the Peclet numbers Pe\mathrm{Pe} as functions of δ\delta and the phase-diagram constant ζ\zeta, but, as in the free-dendrite problem, an additional condition is needed to determine RR and vv separately. Microsolvability theory[30] supplies this through the stability constant

σ=d0Pe R=2Dld0vR2,d0=12(dα+dβ),(54)\sigma= \frac{d_0}{\mathrm{Pe}\,R} = \frac{2D_l d_0}{vR^2}, \qquad d_0 = \frac{1}{2}(d_\alpha+ d_\beta), \tag*{(54)}

which, for given anisotropy strength ϵ1\epsilon_1 of the α\alpha–liquid free energy, takes a value independent of δ\delta and hence of ΔT\Delta T. The PF simulations confirm this prediction: σ≈0.04\sigma\approx0.04 in 3D and σ≈0.14\sigma\approx0.14 in 2D, both essentially constant over the simulated range ΔT∈[20,200] K\Delta T \in[20, 200]\ \mathrm{K} (Figure 5f and Supplementary Figure 4). Combining Eq. (54) with the Ivantsov-type relations yields

R∝d0σδ,v∝Dlσδ2d0,(55)R \propto\frac{d_0}{\sigma\delta}, \qquad v \propto\frac{D_l\sigma\delta^2}{d_0}, \tag*{(55)}

with ζ\zeta-dependent prefactors that differ between 2D and 3D. Since δ∝ΔT\delta\propto\Delta T near TpT_p, Eq. (55) reproduces the observed R∼h∼λ0∝ΔT−1R \sim h \sim\lambda_0 \propto\Delta T^{-1} and v∝ΔT2v \propto\Delta T^2 scalings of Figure 5a–Figure 5d and Supplementary Figure 4–Supplementary Figure 4 without adjustable parameters.

Supplementary Note 5: Supplementary Figures

References

  1. [1]Yu-Guo Guo, Jin-Song Hu, and Li-Jun Wan. Nanostructured materials for electrochemical energy conversion and storage devices. Advanced Materials, 20(15):2878–2887, 2008.
  2. [2]Hui Shao, Yih-Chyng Wu, Zifeng Lin, Pierre-Louis Taberna, and Patrice Simon. Nanoporous carbon for electrochemical capacitive energy storage. Chemical Society Reviews, 49(10):3005–3039, 2020.DOI
  3. [3]Volkmar Zielasek, Birte Jürgens, Christian Schulz, Jürgen Biener, Monika M Biener, Alex V Hamza, and Marcus Bäumer. Gold catalysts: nanoporous gold foams. Angewandte Chemie International Edition, 45(48):8241–8244, 2006.
  4. [4]A Wittstock, V Zielasek, J Biener, CM Friend, and M Bäumer. Nanoporous gold catalysts for selective gas-phase oxidative coupling of methanol at low temperature. Science, 327(5963):319–322, 2010.DOI
  5. [5]Kongcheng Hu, Dongxiao Lan, Xuemei Li, and Shusheng Zhang. Electrochemical dna biosensor based on nanoporous gold electrode and multifunctional encoded dna- au bio bar codes. Analytical chemistry, 80(23):9124–9130, 2008.DOI
  6. [6]Jintao Zhang and Chang Ming Li. Nanoporous metals: fabrication strategies and advanced electrochemical applications in catalysis, sensing and energy systems. Chemical Society Reviews, 41(21):7016–7031, 2012.DOI
  7. [7]AM Hodge, J Biener, JR Hayes, PM Bythrow, CA Volkert, and AV Hamza. Scaling equation for yield strength of nanoporous open-cell foams. Acta Materialia, 55(4):1343–1349, 2007.DOI
  8. [8]Hai-Jun Jin, Lilia Kurmanaeva, Jörg Schmauch, Harald Rösner, Yulia Ivanisenko, and Jörg Weissmüller. Deforming nanoporous metal: Role of lattice coherency. Acta Materialia, 57(9):2665–2672, 2009.DOI
  9. [9]Jonah Erlebacher, Michael J Aziz, Alain Karma, Nikolay Dimitrov, and Karl Sieradzki. Evolution of nanoporosity in dealloying. Nature, 410(6827):450–453, 2001.
  10. [10]JR Hayes, AM Hodge, J Biener, AV Hamza, and Karl Sieradzki. Monolithic nanoporous copper by dealloying Mn–Cu. Journal of Materials Research, 21(10):2611–2616, 2006.DOI
  11. [11]J Snyder, T Fujita, MW Chen, and J Erlebacher. Oxygen reduction in nanoporous metal–ionic liquid composite electrocatalysts. Nature materials, 9(11):904–907, 2010.DOI
  12. [12]Ian McCue, Alain Karma, and Jonah Erlebacher. Pattern formation during electrochemical and liquid metal dealloying. Mrs Bulletin, 43(1):27–34, 2018.DOI
  13. [13]John David Harrison and C Wagner. The attack of solid alloys by liquid metals and salt melts. Acta Metallurgica, 7(11):722–735, 1959.
  14. [14]Takeshi Wada, Kunio Yubuta, Akihisa Inoue, and Hidemi Kato. Dealloying by metallic melt. Materials Letters, 65(7):1076–1078, 2011.DOI
  15. [15]Pierre-Antoine Geslin, Ian McCue, Bernard Gaskey, Jonah Erlebacher, and Alain Karma. Topology-generating interfacial pattern formation during liquid metal dealloying. Nature communications, 6(1):1–8, 2015.DOI
  16. [16]Takeshi Wada, Tetsu Ichitsubo, Kunio Yubuta, Haruhiko Segawa, Hirokazu Yoshida, and Hidemi Kato. Bulk-nanoporous-silicon negative electrode with extremely high cyclability for lithium-ion batteries prepared using a top-down process. Nano letters, 14(8):4505–4510, 2014.DOI
  17. [17]Young Wook Kim, Masashi Tsuda, Takeshi Wada, Kunio Yubuta, Sung Gyoo Kim, and Hidemi Kato. Optimizing niobium dealloying with metallic melt to fabricate porous structure for electrolytic capacitors. Acta Materialia, 84:497–505, 2015.DOI
  18. [18]Ian McCue, Stephen Ryan, Kevin Hemker, Xiandong Xu, Nan Li, Mingwei Chen, and Jonah Erlebacher. Size effects in the mechanical properties of bulk bicontinuous Ta/Cu nanocomposites made by liquid metal dealloying. Advanced Engineering Materials, 18(1):46–50, 2016.DOI
  19. [19]G Greenidge and J Erlebacher. Porous graphite fabricated by liquid metal dealloying of silicon carbide. Carbon, 165:45–54, 2020.DOI
  20. [20]Longhai Lai, Bernard Gaskey, Alyssa Chuang, Jonah Erlebacher, and Alain Karma. Topological control of liquid-metal-dealloyed structures. Nature Communications, 13(1):2918, 2022.DOI
  21. [21]Longhai Lai, Pierre-Antoine Geslin, and Alain Karma. Microstructural pattern formation during liquid metal dealloying: Phase-field simulations and theoretical analyses. Physical Review Materials, 6(9):093803, 2022.DOI
  22. [22]Ruirui Song, Jiuhui Han, Masayuki Okugawa, Rodion Belosludov, Takeshi Wada, Jing Jiang, Daixiu Wei, Akira Kudo, Yuan Tian, Mingwei Chen, et al. Ultrafine nanoporous intermetallic catalysts by high-temperature liquid metal dealloying for electrochemical hydrogen production. Nature Communications, 13(1):5157, 2022.DOI
  23. [23]Zhen Lu, Cheng Li, Jiuhui Han, Fan Zhang, Pan Liu, Hao Wang, Zhili Wang, Chun Cheng, Linghan Chen, Akihiko Hirata, et al. Three-dimensional bicontinuous nanoporous materials by vapor phase dealloying. Nature communications, 9(1):1–7, 2018.
  24. [24]Jiuhui Han, Cheng Li, Zhen Lu, Hao Wang, Zhili Wang, Kentaro Watanabe, and Mingwei Chen. Vapor phase dealloying: A versatile approach for fabricating 3d porous materials. Acta Materialia, 163:161–172, 2019.DOI
  25. [25]Zhen Lu, Fan Zhang, Daixiu Wei, Jiuhui Han, Yanjie Xia, Jing Jiang, Mingwang Zhong, Akihiko Hirata, Kentaro Watanabe, Alain Karma, et al. Vapor phase dealloying kinetics of mnzn alloys. Acta Materialia, 212:116916, 2021.DOI
  26. [26]Wen-Kai Hu, Jun-Chao Shao, Shao-Gang Wang, and Hai-Jun Jin. Evolution of a bicontinuous structure in peritectic melting: The simplest form of dealloying. Physical Review Materials, 3(11):113601, 2019.
  27. [27]Zhongyang Li, Lukas Lührs, Tobias Krekeler, and Jörg Weissmüller. How peritectic melting forms bicontinuous microstructures. Acta Materialia, 289:120917, 2025.
  28. [28]T Muschik, WA Kaysser, and T Hehenkamp. Melting of Cu-In solid solutions at small superheating by droplet formation and liquid film migration. Acta Metallurgica, 37(2):603–613, 1989.DOI
  29. [29]G Boussinot, EA Brener, and DE Temkin. Kinetics of isothermal phase transformations above and below the peritectic temperature: Phase-field simulations. Acta materialia, 58(5):1750–1760, 2010.DOI
  30. [30]A Barbieri and JS Langer. Predictions of dendritic growth rates in the linearized solvability theory. Physical Review A, 39(10):5314, 1989.pubmed.ncbi.nlm.nih.gov/9901096
  31. [31]Tomorr Haxhimali, Alain Karma, Frédéric Gonzales, and Michel Rappaz. Orientation selection in dendritic evolution. Nature materials, 5(8):660–664, 2006.pubmed.ncbi.nlm.nih.gov/16845416
  32. [32]Ilya M Lifshitz and Vitaly V Slyozov. The kinetics of precipitation from supersaturated solid solutions. Journal of physics and chemistry of solids, 19(1-2):35–50, 1961.DOI
  33. [33]C Wagner. Theory of the aging of precipitates by dissolution-reprecipitation (ostwald ripening). Z Elektrochem, 65(7):581–11, 1961.
  34. [34]GP Ivantsov. The temperature field around a spherical, cylindrical, or pointed crystal growing in a cooling solution. In Dokl. Akad. Nauk SSSR, volume 58, pages 567–569, 1947.
  35. [35]Efim A Brener and DE Temkin. Velocity-selection problem for combined motion of melting and solidification fronts. Physical review letters, 94(18):184501, 2005.DOI
  36. [36]Efim A Brener and DE Temkin. Melting of alloys along the inter-phase boundaries in eutectic and peritectic systems. Acta materialia, 55(8):2785–2789, 2007.DOI
  37. [37]R Folch and M Plapp. Quantitative phase-field modeling of two-phase growth. Physical Review E, 72(1):011602, 2005.pubmed.ncbi.nlm.nih.gov/16089974
  38. [38]William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Seminal graphics: pioneering efforts that shaped the field, pages 347–353. 1998.
  39. [39]Alan T Dinsdale. SGTE data for pure elements. CALPHAD, 15(4):317–425, 1991.DOI
  40. [40]Mei Li, Changrong Li, Fuming Wang, and Weijing Zhang. Experimental study and thermodynamic assessment of the Ag–Ti system. CALPHAD, 29(4):269–275, 2005.DOI
  41. [41]Mingwang Zhong, Adriana Eres-Castellanos, Kaihua Ji, Alec I Saville, Brian Rodgers, Dan R Coughlin, John W Gibbs, John D Roehling, Joseph T McKeown, Amy J Clarke, et al. Quantification and prediction of solidification textures under additive manufacturing conditions. Nature Communications, 2025.DOI
  42. [42]Damien Tourret and Alain Karma. Growth competition of columnar dendritic grains: A phase-field study. Acta Materialia, 82:64–83, 2015.DOI
  43. [43]T Ihle and H Müller-Krumbhaar. Fractal and compact growth morphologies in phase transitions with diffusion transport. Physical Review E, 49(4):2972, 1994.pubmed.ncbi.nlm.nih.gov/9961564
  44. [44]Silvère Akamatsu, Gabriel Faivre, and Thomas Ihle. Symmetry-broken double fingers and seaweed patterns in thin-film directional solidification of a nonfaceted cubic crystal. Physical Review E, 51(5):4751, 1995.pubmed.ncbi.nlm.nih.gov/9963188
  45. [45]Nathalie Limodin, Luc Salvo, Elodie Boller, Michel Suéry, M Felberbaum, Sylvain Gaillègue, and Kamel Madi. In situ and real-time 3-d microtomography investigation of dendritic solidification in an al–10 wt.% cu alloy. Acta Materialia, 57(7):2300–2310, 2009.DOI
  46. [46]Pierre-Antoine Geslin, Mickael Buchet, Takeshi Wada, and Hidemi Kato. Phase-field investigation of the coarsening of porous structures by surface diffusion. Physical Review Materials, 3(8):083401, 2019.DOI
  47. [47]Morgane Mokhtari, Christophe Le Bourlot, Jérôme Adrien, Anne Bonnin, Wolfgang Ludwig, Pierre-Antoine Geslin, Takeshi Wada, Jannick Duchet-Rumeau, Hidemi Kato, and Eric Maire. In situ observation of liquid metal dealloying and etching of porous fecr by x-ray tomography and x-ray diffraction. Materialia, 18:101125, 2021.DOI

Paper details

vlnc:W_dcr4v86s
Version
v1 · latest
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.

  • Physics and Astronomy

Claude Opus 4.7 (bare)

View statement
Looking for related papers…
Contents