Reproduce “Physics-Informed Transformer for Real-Time High-Fidelity Topology Optimization” on OpenResearch

About this paper

Aaron Lutheran, Srijan Das, Alireza Tabarraei

Researchers at the University of North Carolina at Charlotte developed a physics-informed transformer for real-time topology optimization, which learns a direct mapping from physical inputs to optimized structures. The method achieved a computational speedup exceeding three orders of magnitude for dynamic problems, delivering high-fidelity designs for static scenarios while effectively balancing efficiency and structural accuracy in ML-based approaches.

Key insights

• Reframing topology optimization as a global operator-learning problem using transformers effectively addresses the trade-off between computational speed and structural fidelity in ML-based approaches.
• The global self-attention mechanisms of Vision Transformers are well-suited for modeling long-range mechanical interactions in structural design, outperforming local convolutional architectures when an optimal patch size (e.g., P=4) is chosen.
• Differentiable auxiliary losses, particularly the novel floating-material loss, are crucial for generating physically realistic and manufacturable designs within a gradient-based learning framework, enabling explicit enforcement of structural connectivity.

Summary by alphaXiv.