Modified augmented Lagrangian preconditioning for mixed-dimensional beam-solid coupling

Technical Summary: Modified Augmented Lagrangian Preconditioning for Mixed-Dimensional Beam-Solid Coupling

Problem Statement
The paper addresses the numerical challenges associated with the mixed-dimensional coupling of three-dimensional (3D) solid bodies and embedded one-dimensional (1D) torsion-free Kirchhoff–Love beams. This modeling approach is essential for simulating fiber-reinforced materials, such as short-fiber composites or biological tissues, where fibers are significantly slender compared to the solid matrix. The coupling is enforced via Lagrange multipliers to ensure exact positional constraints between the beam centerline and the surrounding solid.

This formulation leads to a linear system of equations with an indefinite saddle-point structure. The system presents specific difficulties:

1. Indefiniteness: The presence of a zero block on the diagonal (associated with the Lagrange multipliers) prevents the use of standard positive-definite solvers.
2. Singular Sub-problems: The beam sub-problem, when considered in isolation (as required by block preconditioners), is a pure Neumann problem due to the lack of explicit Dirichlet boundary conditions on the embedded fibers. This renders the beam stiffness matrix singular.
3. Parameter Sensitivity: The conditioning of the system is heavily influenced by modeling parameters, including the stiffness ratio between beam and solid ($E = E_B/E_S$), the beam-to-solid volume ratio ($V$), and the mesh size ratio.
4. Limitations of Existing Methods: Previous work by the authors utilized penalty regularization, which reduces system size but introduces ill-conditioning and variational inconsistency. Conversely, direct Lagrange multiplier approaches require specialized preconditioning strategies that are not yet fully explored for this specific mixed-dimensional, vector-valued elasticity context.

Methodology
The authors propose a Modified Augmented Lagrangian (MAL) block preconditioning strategy to solve the resulting saddle-point system. The methodology proceeds as follows:

1. Augmented Formulation: The original Lagrange multiplier system is reformulated into an augmented Lagrangian system. This involves adding penalty-like terms ($\epsilon M^T W^{-1} M$ and $\epsilon D^T W^{-1} D$) to the solid and beam stiffness blocks. Crucially, this augmentation:

   • Preserves the exact enforcement of coupling constraints (unlike pure penalty methods).
   • Regularizes the singular beam sub-problem, making the beam stiffness block invertible even for small penalty parameters.
   • Maintains the saddle-point structure of the overall system.

2. Block Preconditioner Construction: Starting from an ideal augmented Lagrangian preconditioner, the authors derive practical block-triangular variants. These preconditioners allow for the independent treatment of the solid, beam, and Schur complement blocks:

   • Solid Block: Preconditioned using Algebraic Multigrid (AMG) methods (specifically Smoothed Aggregation and Plain Aggregation).
   • Beam Block: Since the augmented beam matrix is block-diagonal (for independent fibers) and invertible, it is solved using direct methods (LU factorization).
   • Schur Complement: The core of the method involves approximating the augmented Schur complement. The authors introduce three distinct variants:
     • Variant I: Approximates the Schur complement using sparse approximate inverses (SPAI) for the beam block and a diagonal approximation for the solid block, assuming a small penalty parameter ($\epsilon \ll E_B$).
     • Variant II: Uses a diagonal approximation based on the scaling matrix $W$ (specifically $\hat{S} \approx -\frac{1}{\epsilon}W$), requiring a carefully tuned penalty parameter to cluster eigenvalues.
     • Variant III: Introduces separate penalty parameters for the solid ($\epsilon_S$) and beam ($\epsilon_B$) contributions. By setting $\epsilon_S \to 0$ and tuning $\epsilon_B$, this variant aims to decouple the scaling issues, utilizing a diagonal Schur complement approximation.

3. Numerical Implementation: The linear systems are solved using GMRES. The implementation leverages the Trilinos and 4C software frameworks, utilizing distributed-memory parallelism (MPI).

Key Contributions

• Novel Preconditioning Framework: The paper presents the first application of modified augmented Lagrangian block preconditioners specifically tailored to mixed-dimensional beam-solid coupling with Lagrange multipliers.
• Handling of Singular Sub-problems: The augmentation strategy effectively regularizes the singular beam sub-problem without resorting to complex null-space projection techniques or deflation methods, allowing for the use of standard direct solvers on the beam block.
• Schur Complement Variants: The systematic derivation and comparison of three Schur complement approximation strategies (I, II, and III), including the introduction of independent penalty parameters for solid and beam domains (Variant III).
• Scalability Analysis: Comprehensive evaluation of strong and weak scaling on distributed-memory clusters for problems with millions of degrees of freedom.

Results
Numerical experiments on representative volume elements (RVEs) and a hybrid composite plate application yielded the following findings:

• Robustness: The proposed preconditioners demonstrate robustness with respect to the stiffness ratio ($E$), volume ratio ($V$), and mesh refinement. Variant III, in particular, showed near parameter-robustness for moderate to high stiffness contrasts.
• Mesh Independence: The methods achieved near mesh-independent iteration counts (typically between 10 and 50 iterations for RVEs, and around 20 for the composite plate) across different discretization densities.
• Scalability: The methods exhibited favorable strong and weak scaling on parallel clusters (up to 512 processors). Variant II and III showed superior parallel efficiency (up to 97% weak scaling efficiency for setup) compared to Variant I, primarily due to the computational cost of the Schur complement approximation in Variant I.
• Engineering Application: In the application to a hybrid composite sandwich plate, the naive ILU preconditioner failed to converge. In contrast, the proposed MAL preconditioners (especially Variant IIIc) achieved convergence with low iteration counts and significantly reduced solution times (up to 4x faster than the worst-performing variant) compared to standard approaches.

Significance
The paper claims that the proposed modified augmented Lagrangian block preconditioners provide a robust and efficient alternative to both pure penalty-based formulations and standard Lagrange multiplier approaches for large-scale simulations of mixed-dimensional models in solid and structural mechanics. By enabling exact constraint enforcement while regularizing singular sub-problems, the method facilitates the simulation of complex fiber-reinforced structures with high stiffness contrasts and varying volume fractions. The authors conclude that this approach is particularly suitable for engineering applications involving embedded fibers, offering a viable path toward scalable, high-fidelity simulations of composite materials. Future work is noted to extend this framework to more complex beam formulations (e.g., Simo–Reissner) that include torsion and shear effects.