Scalable Stress-Constrained Topology Optimization: A Super-resolution and Simultaneous Approach Based on Machine Learning

Introduction

Topology optimization (TO) is a powerful tool for designing lightweight structures, but stress‑constrained variants quickly become computationally prohibitive as the number of degrees of freedom (DoF) grows. A recent pre‑print proposes a Simultaneous Neural Topology Optimization (SNTO) framework that replaces traditional finite‑element solvers with custom neural networks (NNs) to represent the state, adjoint, and design fields. By training these networks jointly through a single optimization loop, the method achieves super‑resolution designs while scaling to millions of collocation points on multi‑GPU systems.

Neural Representation of TO Variables

SNTO encodes the physical fields (displacement, stress, design density) as outputs of separate NNs whose parameters are updated together. The loss function combines three components:

1. Design objective (e.g., compliance minimization),
2. Aggregated stress constraints, and
3. Potential energy functional to enforce equilibrium.

Automatic differentiation supplies exact gradients for all terms, eliminating the need for separate sensitivity analyses. The network architecture is tuned to capture sharp material boundaries and high stress gradients—features that are essential for accurate stress‑constrained designs.

Hardware‑Aware Partitioning for Scalability

To handle problems with up to eight million collocation points, the authors introduce a domain‑decomposition strategy that is native to the neural formulation. The design domain is split into subdomains that are distributed across multiple GPUs. Within each GPU, subdomains are further sliced into smaller regions; loss gradients are computed sequentially for these slices and accumulated before updating the NN weights. This approach:

• Keeps peak GPU memory usage low,
• Guarantees continuity across subdomain interfaces without adding extra penalty terms, and
• Leverages parallelism to reduce wall‑clock time.

Validation and Performance

The framework is evaluated on several 2‑D benchmarks with stress constraints. Compared to a conventional finite‑element based optimizer, SNTO delivers:

• Sharper material boundaries and reduced gray‑scale regions, indicating higher design fidelity,
• Lower volume fractions for comparable compliance, showing more efficient material usage,
• Super‑resolution outputs that exceed the spatial resolution of the underlying collocation grid.

For large‑scale 3‑D cases, the method scales smoothly across up to eight GPUs, demonstrating practical feasibility for industrial‑size problems.

Practical Takeaways

• Unified training loop: By treating state, adjoint, and design variables as learnable functions, SNTO simplifies the optimization pipeline and exploits modern deep‑learning toolkits.
• Memory‑efficient partitioning: The hierarchical slicing strategy makes it possible to train on commodity GPU clusters without sacrificing continuity.
• Design quality: The neural representation naturally yields high‑contrast, low‑gray designs, which can reduce post‑processing effort in manufacturing.

Conclusion

SNTO illustrates how machine‑learning techniques can be woven directly into the core of topology optimization, turning a traditionally iterative, solver‑heavy process into a gradient‑driven learning problem. Its ability to handle stress constraints at scale, while delivering super‑resolution designs, points toward a new class of TO tools that are both computationally tractable and capable of producing high‑quality engineering solutions.