Fail-safe topology optimization using artificial neural networks

Accounting for fail-safety in topology optimization requires the consideration of a large number of failure scenarios. For each scenario, the structural responses must be evaluated, rendering explicit fail-safe topology optimization computationally extremely demanding. In recent years, artificial neural networks (ANNs) have gained increasing attention in topology optimization as a means to overcome existing limitations and advance the state of the art. To explore the potential and limitations of ANN-based fail-safe topology optimization, this work proposes a novel framework for generating fail-safe designs. Within this framework, fail-safe designs are derived from designs optimized for minimal compliance using an ANN. In addition, a post-processing step is introduced to address common problems associated with topology-optimized designs predicted by ANNs. Compared to the traditional approach, the proposed framework reduces the computing time required to generate fail-safe designs by 84.98%. However, due to the high initial cost of data generation and model training, achieving a net computational benefit would require very large number of applications. Compared with deterministic designs, the generated fail-safe designs achieve a median reduction in damage compliance of 96.02%. However, compared with the traditional fail-safe approach, the damage compliance is 85.32% higher. Moreover, the results indicate limited versatility, with performance degrading as problem formulations deviate increasingly from the training distribution.