A study by Muhammad Habibur Rahman, Khaled M. Furati, and Abdullah Shah addresses a persistent computational bottleneck in this field. The authors developed a numerical simulation of a two-dimensional phase-field model for dendritic solidification that uses a fixed, algebraic grid stretching transformation to concentrate computational resolution where it matters most: near the evolving solid-liquid interface. Their approach, formulated in curvilinear coordinates with a structured finite-difference discretization, offers a low-complexity alternative to adaptive mesh refinement, and the paper provides an unusually candid assessment of both its advantages and its limitations.

The phase-field method has become one of the most widely used frameworks for simulating solidification because it represents the solid-liquid interface as a diffuse transition layer, avoiding the need for explicit front tracking. The researchers adopted the classical Kobayashi model, which couples a dimensionless phase-field variable, an order parameter that smoothly transitions from zero in the liquid to one in the solid, with a normalized temperature field. Anisotropy enters through an orientation-dependent interface-width parameter, controlled by an anisotropy strength and a mode number, so that growth is favored along preferred crystallographic directions. The latent heat released at the moving interface feeds back into the temperature field, which in turn modifies the thermodynamic driving force for solidification.

The difficulty is resolution. Accurate phase-field simulations require fine spatial sampling across the diffuse interface and the surrounding thermal boundary layer, yet on a uniform grid the entire domain must be refined even though the strongest gradients remain localized near the interface. Adaptive mesh refinement can handle moving interfaces effectively, but it introduces complexity through dynamic regridding, interpolation, and data transfer. The authors instead evaluated a fixed, separable algebraic coordinate transformation, chosen in advance based on the expected initial location of the interface, which clusters grid points in the region of anticipated activity while preserving a structured mesh topology and avoiding remeshing entirely.

Transforming the governing equations to curvilinear coordinates is not a trivial bookkeeping exercise. The chain rule introduces metric terms, including the Jacobian of the mapping and its derivatives, that modify every spatial operator. The authors derived the transformed anisotropic phase-field equation and the transformed temperature equation in full, and they also derived a metric-dependent stability estimate for the explicit Forward Euler time integration. Because the computational grid is uniform in computational coordinates but nonuniform in physical space, the admissible time step is governed by the smallest effective physical spacing in the clustered region, so stronger clustering improves local sampling while simultaneously increasing the number of required time steps.

Verification was carried out in three complementary stages. First, a one-dimensional traveling-front benchmark, which retains a steep diffuse-interface profile and admits an exact hyperbolic-tangent reference solution, was used to test the stretched-grid phase-field operator. Under coupled space-time refinement, in which the time step scales with the square of the grid spacing, the observed fully discrete convergence orders approached two on both uniform and stretched grids. Notably, at every tested resolution the stretched grid produced a smaller relative physical-domain error than the uniform grid; at the finest tested level of 3200 computational intervals, the errors were approximately 4.3 by 10 to the minus 6 for the stretched grid versus 5.3 by 10 to the minus 6 for the uniform grid.

The cost picture, however, is more sobering, and the authors are explicit about it. Both grid formulations first satisfied the prescribed error target of 10 to the minus 5 at the same resolution, and the measured median CPU times at that level were 2.8758 seconds for the uniform grid implementation and 18.592 seconds for the stretched grid implementation. The study therefore demonstrates an accuracy benefit at fixed resolution, but not a computational-time advantage, a conclusion the authors confine to their one-dimensional implementation and fixed mapping. This honesty distinguishes the work from claims that nonuniform gridding automatically saves computation.

Two further verification tests examined the two-dimensional machinery. An isotropic manufactured-solution test, with manufactured phase and temperature fields on a periodic-like square domain, confirmed that the transformed diffusion operators, the phase-temperature coupling, the homogeneous Neumann boundary treatment, and the Forward Euler time discretization converge at rates approaching two under coupled refinement. A separate analytical-field test with nonzero anisotropy assessed the physical interface orientation, the spatial derivatives of the anisotropy-dependent coefficients, and the assembled anisotropic spatial operator, again with finest-grid orders near two. The authors note that this last test is restricted to the interior discretization for smooth fields with nonvanishing gradients; the orientation treatment at zero-gradient points and the anisotropic boundary closure were not independently verified.

With the numerics verified, the authors turned to dendritic growth simulations on a square nondimensional domain, using parameter values adopted from the Kobayashi model literature. Three initial configurations were explored: a diffuse circular seed at the domain center, a thin solid layer along the lower boundary, and a solid layer centered in the middle of the domain, each paired with a grid stretching tailored to the expected interface location. All three produced characteristic fourfold dendritic morphologies consistent with the chosen anisotropy mode number of four, with side branches and evolving temperature fields driven by latent-heat release. A grid-sensitivity study of the central-growth case showed that the coarsest stretched grid differed from the finest tested reference by about 11.1 percent in tip-velocity history and 9.2 percent in final tip growth, differences that fell to roughly 3 percent on an intermediate grid, indicating that under-resolution produces a noticeable but diminishing effect on calculated kinetics.

A parametric study then illustrated how morphology responds to three dimensionless parameters. Increasing the latent-heat coupling parameter transformed the crystal from nearly circular growth to a diamond-like fourfold shape and finally to a branching dendritic structure with thinner, more defined arms. Increasing the relaxation-time parameter shortened and thickened the branches, producing more compact, rounded structures, because the interface responds more slowly to gradients. Varying the interface-width parameter changed the sharpness of the depiction, with narrower interfaces yielding more compact crystals and wider ones producing smoother, more extended growth. The authors caution that in one configuration the diffuse interface was only marginally resolved in the coarse outer region, so fine morphological features there should be interpreted qualitatively.

The central limitation is inherent to the method: the fixed algebraic mapping is selected from the initial interface configuration, so its local resolution advantage diminishes as the interface moves away from the clustered region. Unlike adaptive mesh refinement, the mapping does not dynamically maintain a prescribed number of grid points across the diffuse interface. Quantitative benchmarking of tip position, tip velocity, interface curvature, and thermal-field errors against a highly resolved two-dimensional reference solution also remains outside the present scope. The authors additionally note that all reported quantities are nondimensional, with no material-specific calibration, so the reported lengths and times do not have a unique conversion to physical units.

Within those boundaries, the study makes a carefully scoped contribution: a transformed anisotropic formulation, a stability estimate that accounts for the metric terms of the stretching, and a numerical assessment of static point clustering as the interface migrates. The authors explicitly make no claim of adaptive refinement or a new thermodynamic phase-field model, and they point toward future work on dynamically updated mappings, adaptive refinement, and more efficient time integrators. For researchers simulating dendritic solidification, the message is nuanced: fixed algebraic grid stretching can buy accuracy at a fixed number of grid points, but it does not, in this implementation, buy speed, and its benefits fade as the dendrite grows beyond the region where the grid was refined.

Subject of Research: Numerical simulation of dendritic crystal growth using a phase-field model with fixed algebraic grid stretching

Article Title: Numerical simulation of a phase-field model for dendritic crystal growth using algebraic grid stretching

Article References: Rahman, M. H., Furati, K. M., & Shah, A. (2026). Numerical simulation of a phase-field model for dendritic crystal growth using algebraic grid stretching. Results in Engineering, 32, Article 113719. https://doi.org/10.1016/j.rineng.2026.113719

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113719

Keywords: phase-field model, dendritic crystal growth, grid stretching, curvilinear coordinates, solidification, Kobayashi model, finite differences, adaptive mesh refinement, numerical verification, anisotropy, latent heat, computational cost