Thermal Optimization with Low-Rank Tensor Representation of Operator Learning Models

Thermal optimization is critical in 3D-IC design but computationally prohibitive with traditional methods. We propose a low-rank tensor representation of DeepOHeat that enables thermal simulation on high-resolution problems previously intractable for operator learning frameworks. By decomposing basis functions along coordinate axes, our approach makes training possible on large 3D thermal problems where standard methods fail due to memory constraints. When integrated as a surrogate model in thermal optimization for floorplan, our method produces designs comparable to high-fidelity solvers while reducing optimization time from 4.94 hours to 0.05 hours.