Adversarial Agents on Topology Optimization: Understanding the Fragility and Robustness of Deep Learning-based and Physics-Based Design Models under Adversarial Perturbation

Topology optimization is widely used in automated cyber-manufacturing to synthesize lightweight structures. To overcome the computational bottleneck of iterative finite-element solvers, deep learning surrogates have been adopted to generate designs instantly. However, connected cyber-manufacturing systems require physical trustworthiness against input anomalies and geometry corruption. This paper investigates whether deep learning-based topology optimization models are vulnerable to small, bounded input perturbations, and whether physics-in-the-loop recovery can mitigate this vulnerability. The authors formulate a mechanics-grounded reliability evaluation framework featuring an adversarial agent that targets generative design models. They examine a strictly non-intrusive threat model where bounded perturbations are introduced exclusively to the initial-density channel, while physical boundary conditions, solver routines, and network architectures remain unchanged. They evaluate three surrogate families—U-Net, convolutional networks, and generative adversarial models—across varying physics-gradient conditioning depths.

Evaluating surrogate models reveals that bounded initialization noise can cause catastrophic mechanical failure, increasing structural compliance by multiple orders of magnitude through severed load paths and disconnected supports. Standard machine learning metrics such as pixel accuracy or mean squared error fail to capture these failures, whereas mechanics-grounded compliance evaluation exposes them clearly. Furthermore, incorporating richer physics-gradient conditioning in the deep learning surrogates does not guarantee monotonic robustness across surrogate families. When an adversarial initialization degrades a deep learning surrogate, initializing the classical SIMP optimizer with the perturbed topology mitigates performance degradation, successfully restoring structural compliance to near-baseline levels with high probability.

These findings demonstrate that learned surrogates should serve as physics-verified initializers instead of replacing physics-based solvers entirely in a resilient cyber-manufacturing system. While deep learning surrogates offer rapid generation, relying on them without physical verification introduces severe structural risks under input perturbations. Conversely, combining learned initializers with classical physics-based optimization provides both computational speed and mechanical reliability. The proposed adversarial agent also establishes a foundational framework for training generative design agents that are robust against noise and targeted perturbations.