Abstract
We propose a new framework for topology optimization based on the boundary element discretization and kernel-independent fast multipole method (KIFMM). The boundary value problem for the considered partial differential equation is reformulated as a surface integral equation and is solved on the domain boundary. Volume solution at selected points is found via surface integrals. At every iteration of the optimization process, the new boundary is extracted as a level set of a topological derivative. Both surface and volume solutions are accelerated using KIFMM. The obtained technique is highly universal, fully parallelized, it allows achieving asymptotically the best performance with the optimization iteration complexity proportional to a number of surface discretization elements. Moreover, our approach is free of the artifacts that are inherent for finite element optimization techniques, such as "checkerboard" instability. The performance of the approach is showcased on few illustrative examples.
Introduction
Topology optimization is a computational technique used in engineering to determine the most efficient shape for a structure—such as a bridge bracket or an aerospace component—by minimizing material usage while maintaining structural integrity. Traditionally, this process relies on the Finite Element Method (FEM), which discretizes a 3D volume into small elements like cubes or tetrahedrons. However, FEM-based optimization often suffers from numerical artifacts, most notably "checkerboard" instabilities where solid and void cells alternate in a non-physical pattern. Furthermore, as designs become more complex, especially with the advent of additive manufacturing, the computational cost of FEM scales poorly.
This paper presents a novel framework that addresses these issues by employing the Boundary Element Method (BEM) accelerated by a Kernel-Independent Fast Multipole Method (KIFMM). This combination enables high-fidelity 3D topology optimization with near-linear parallel scalability.
Methodology
The proposed framework shifts the focus from the volume to the boundary. By reformulating the partial differential equations of elasticity into surface integral equations, the Boundary Element Method (BEM) only requires discretization of the object's surface. This reduction in dimensionality from 3D to 2D significantly decreases the number of elements needed for a given resolution.
To overcome the $O(N^2)$ complexity typically associated with BEM's dense matrices, the researchers integrated the Kernel-Independent Fast Multipole Method (KIFMM). This "engine" reduces the computational complexity to approximately $O(N)$, where $N$ is the number of surface elements. The kernel-independent nature of this method ensures that the framework can be adapted to various physics problems, such as heat transfer or electromagnetics, without extensive code modification.
The optimization process evolves the structure's boundary as a level set of a topological derivative (TD). At each iteration, the TD is calculated at every point in the volume to determine how material changes would affect the total energy. The entire pipeline is parallelized using the PETSc library for linear algebra and the PVFMM library for fast multipole calculations, allowing the workload to be distributed across thousands of CPU cores.
Results and Discussion
The framework was evaluated using classical structural engineering benchmarks, specifically 3D cantilever support examples. The initial models were discretized with approximately 64³ voxels, resulting in 0.79 million volume degrees of freedom (DOF) and 4.72 million surface DOFs.
A primary advantage of this BEM-based approach is the natural elimination of "checkerboard" instabilities. Because BEM treats the material as a continuum rather than a collection of independent voxels, the resulting topologies are inherently smoother and more physically consistent.
Performance benchmarks demonstrated that a single optimization iteration for a complex 3D model takes roughly 15 minutes on 64 cores. The volume computation part of the process was particularly fast, processing approximately 2,000 points per second per core. The implementation showed nearly linear scalability up to 64 cores, and given that the underlying PVFMM library is known to scale up to 20,000 cores, the method is well-suited for massive, industrial-scale problems.
Conclusion and Future Work
The combination of BEM and KIFMM provides a powerful and scalable alternative to traditional FEM for 3D topology optimization. By focusing on surface discretization and leveraging fast multipole methods, the framework achieves high performance while avoiding common numerical artifacts.
Future work will address current limitations, such as improving the accuracy of topological derivatives near boundaries through local extrapolation or surface mesh refinement. Additionally, the researchers plan to parallelize the initial meshing and final post-processing steps, which are currently serial bottlenecks. Developing more efficient preconditioners for the iterative GMRES solver and extending the method to handle non-homogeneous or composite materials are also key priorities for future development.
Summary updated