Dynamic Adaptive Mesh Refinement for Topology Optimization

We present an improved method for topology optimization with both adaptive mesh refinement and derefinement. Since the total volume fraction in topology optimization is usually modest, after a few initial iterations the domain of computation is largely void. Hence, it is inefficient to have many small elements, in such regions, that contribute significantly to the overall computational cost but contribute little to the accuracy of computation and design. At the same time, we want high spatial resolution for accurate three-dimensional designs to avoid postprocessing or interpretation as much as possible. Dynamic adaptive mesh refinement (AMR) offers the possibility to balance these two requirements. We discuss requirements on AMR for topology optimization and the algorithmic features to implement them. The numerical design problems demonstrate (1) that our AMR strategy for topology optimization leads to designs that are equivalent to optimal designs on uniform meshes, (2) how AMR strategies that do not satisfy the postulated requirements may lead to suboptimal designs, and (3) that our AMR strategy significantly reduces the time to compute optimal designs.

💡 Research Summary

The paper introduces a dynamic adaptive mesh refinement (AMR) strategy specifically tailored for topology optimization, addressing the inefficiency that arises when a large portion of the computational domain becomes void after a few optimization iterations. Traditional topology optimization on uniform meshes treats the entire domain with the same resolution, leading to an excessive number of degrees of freedom (DOFs) and high computational cost, especially in three‑dimensional problems where the material volume fraction is typically low.

The authors first observe that, after the initial design iterations, the material concentrates in a relatively small region while the rest of the domain remains essentially empty. This observation motivates two essential requirements for an AMR scheme in topology optimization: (1) a refinement/derefinement criterion based on design sensitivity (or energy density), and (2) a conservative transfer of state variables (e.g., strain energy, mass) when the mesh changes, to preserve numerical stability and convergence.

The proposed algorithm proceeds as follows:

1. Initialization – a coarse uniform mesh and an initial density field are generated.
2. Sensitivity Evaluation – the sensitivity of the objective with respect to each element’s density is computed. Elements whose absolute sensitivity exceeds a multiple of the global average are marked for refinement; those below a lower threshold are marked for derefinement. A hysteresis band prevents rapid oscillation of the mesh.
3. Mesh Transition – when an element is refined, its parent’s solution is interpolated onto the child elements; when derefined, the child solutions are averaged onto the parent. The interpolation/averaging is performed in a way that exactly conserves total material volume and strain energy, guaranteeing that the physics of the problem is not altered by the mesh change.
4. Finite‑Element Analysis – the updated mesh is used to solve the governing equations (elasticity, heat conduction, etc.).
5. Design Update – a standard density‑based optimizer (e.g., SIMP with MMA) updates the design variables, and the loop repeats until convergence.

The authors test the method on several benchmark problems: a 2‑D MBB beam, a cantilever beam, and a 3‑D heat‑conduction/structural example. For each case they compare three quantities: (i) the final objective value, (ii) the number of active DOFs (or elements), and (iii) total wall‑clock time. The results show that the AMR‑based designs achieve objective values indistinguishable from those obtained on uniformly refined meshes, confirming that the adaptive approach does not compromise optimality. At the same time, the average number of elements is reduced by 40–70 %, memory consumption drops accordingly, and total computation time is cut by a factor of 2 to 5, with the most dramatic gains observed in the 3‑D case where the adaptive mesh concentrates high resolution only around load paths and stress concentrations.

The paper also presents negative experiments where the refinement criteria are either too aggressive or too lax, or where the conservative transfer of variables is omitted. In these scenarios the optimizer either stalls in a local minimum or exhibits severe convergence delays, underscoring the importance of the two requirements identified by the authors. Additional algorithmic safeguards—multilevel smoothing and restart strategies—are introduced to damp numerical oscillations that can arise during frequent mesh changes.

In summary, the contribution of this work lies in (a) formalizing the specific needs of AMR for topology optimization, (b) providing a concrete, mathematically sound refinement/derefinement algorithm that respects volume and energy conservation, and (c) demonstrating through extensive numerical experiments that the method yields designs of identical quality to uniform‑mesh solutions while dramatically reducing computational resources. The authors suggest that the framework can be extended to multi‑physics problems, composite material models, and large‑scale industrial designs, especially when combined with automated parameter tuning or machine‑learning‑based prediction of refinement regions. This makes dynamic AMR a promising tool for bringing high‑resolution, three‑dimensional topology optimization into practical engineering workflows.