SELTO: Sample-Efficient Learned Topology Optimization
This work addresses the problem of inefficient and unreliable deep learning methods in topology optimization for researchers and engineers, though it is incremental in nature.
The paper tackled the lack of sample efficiency and physical correctness in deep learning for topology optimization by developing physics-based preprocessing and equivariant networks, resulting in drastic improvements in sample efficiency and prediction accuracy. It also introduced the first two public datasets for topology optimization to enhance comparability and future progress.
Recent developments in Deep Learning (DL) suggest a vast potential for Topology Optimization (TO). However, while there are some promising attempts, the subfield still lacks a firm footing regarding basic methods and datasets. We aim to address both points. First, we explore physics-based preprocessing and equivariant networks to create sample-efficient components for TO DL pipelines. We evaluate them in a large-scale ablation study using end-to-end supervised training. The results demonstrate a drastic improvement in sample efficiency and the predictions' physical correctness. Second, to improve comparability and future progress, we publish the two first TO datasets containing problems and corresponding ground truth solutions.