Augmented Lagrangian preconditioning for a simplified Ericksen--Leslie model of nematic liquid crystals

Technical Summary: Augmented Lagrangian Preconditioning for a Simplified Ericksen–Leslie Model

Problem Statement
The numerical solution of the simplified Ericksen–Leslie model for nematic liquid crystals presents significant challenges due to the strong coupling between the fluid flow (velocity $u$) and the director field ($n$), alongside the simultaneous enforcement of two constraints: incompressibility ($\nabla \cdot u = 0$) and the pointwise unit-length condition ($|n| = 1$). While Lagrange multiplier formulations avoid the ill-conditioning associated with small Ginzburg–Landau penalty parameters, they result in Newton systems with a complex double saddle-point structure. Specifically, the linearized system at each time step is a $4 \times 4$ block matrix coupling the velocity–director variables with the pressure–multiplier variables. Designing robust, scalable iterative solvers for these systems remains a critical computational hurdle.

Methodology
The authors propose a fully discrete augmented Lagrangian (AL) approach combined with a specialized block preconditioner.

1. Formulation: The model is reformulated using a scalar Lagrange multiplier $q$ to enforce the unit-length constraint, replacing the cubic nonlinearity $-|\nabla n|^2 n$ with a bilinear term $qn$. The augmented Lagrangian functional introduces quadratic penalty terms for both the incompressibility and unit-length constraints, controlled by parameters $\gamma_u$ and $\gamma_n$. Crucially, the constraints are retained in multiplier form rather than being replaced by pure penalties, ensuring accuracy while improving the spectral properties of the Schur complements.
2. Discretization: The system is discretized in space using Taylor–Hood finite elements ($P_2$ for velocity and director, $P_1$ for pressure and multiplier) and in time using the backward Euler method. The resulting nonlinear algebraic system is linearized via Newton's method.
3. Preconditioning Strategy: A block preconditioner is developed for the resulting saddle-point system. The unknowns are grouped into a primal block (velocity and director increments) and a dual block (pressure and multiplier increments).
   • The coupled velocity–director block is approximated by its block-diagonal components.
   • This leads to decoupled approximations of the two constraint Schur complements:
     • Pressure Schur Complement ($S_p$): Approximated using a generalized Stokes principal part combined with the Sherman–Morrison–Woodbury identity, yielding a form involving the pressure mass matrix and an inverse Laplacian term scaled by the augmentation parameter.
     • Director-Multiplier Schur Complement ($S_q$): Approximated by recognizing the zero-order nature of the constraint operator. The approximation is dominated by the penalty term, resulting in a scaled mass matrix inverse for the multiplier space.

Key Contributions

1. Augmented Lagrangian Formulation: The authors formulate a fully discrete AL system for the time-dependent simplified Ericksen–Leslie equations where both constraints are enforced via multipliers, augmented by penalty terms to improve conditioning.
2. Block Preconditioner Derivation: A novel block-diagonal preconditioner is derived that separates the coupled velocity–director dynamics from the constraint enforcement. The authors construct physically scaled approximations for the pressure and director-multiplier Schur complements, explicitly accounting for the distinct algebraic structures of the divergence and unit-length constraints.
3. Comprehensive Numerical Assessment: The paper provides a rigorous assessment of the method's accuracy and solver robustness through manufactured solutions, energy benchmarks, and various parameter studies (mesh refinement, time-step variation, viscosity changes, and complex geometries).

Results

• Accuracy: Manufactured solution tests confirm the expected spatial accuracy (approximately third-order $L_2$ decay for $P_2$ elements) and first-order temporal convergence for the primary variables (velocity, director, pressure). The multiplier error reaches a spatial-error floor on fixed meshes, consistent with the algebraic nature of the constraint.
• Solver Performance: The outer FGMRES iteration counts are nearly mesh-independent and remain stable under variations in time-step size and fluid viscosity.
• Parameter Sensitivity: Iteration counts improve as the augmentation parameters ($\gamma_u, \gamma_n$) increase. Spectral analysis of the preconditioned Schur blocks shows that the multiplier block's condition number decreases monotonically with increasing $\gamma$, while the pressure block remains uniformly bounded.
• Physical Consistency: A smooth benchmark problem demonstrates monotone decay of the computed total energy, consistent with the continuous energy dissipation law of the Ericksen–Leslie model.
• Complexity: The method successfully handles curved domains (unit disk) and complex initial conditions involving topological defects (two-defect configuration) without modification to the preconditioner.

Significance and Claims
The paper claims that the proposed augmented Lagrangian block preconditioner effectively addresses the ill-conditioning associated with the unit-length constraint in dynamic liquid crystal simulations. By treating the incompressibility and director-length constraints simultaneously within an AL framework, the method achieves a robust solver that is largely independent of mesh size, time step, and viscosity. The authors emphasize that the preconditioner's design relies on the distinct spectral properties of the two constraints, allowing for efficient decoupling.

The work is presented as a significant step toward solving fully discrete, monolithic Ericksen–Leslie systems. The authors modestly note that their current results are limited to the incompressible regime and specific parameter ranges. They identify the extension to compressible nematic flows (where density variations are significant) as a natural but challenging future direction, requiring a reassessment of the saddle-point structure. The paper does not claim universal optimality but demonstrates that the proposed strategy provides stable and efficient convergence for the tested simplified model.