Isa-PIPN: an interconnect-structure-aware physics-informed PointNet method for accurate and efficient capacitance extraction

This paper present an interconnect-structure-aware physics-informed PointNet method, named Isa-PIPN, for fast and accurate capacitance predictions. Once the Isa-PIPN is trained, it can accurately predict the capacitance of multiple sets of interconnect structures with a variable number of conductors, and moreover, the inference time of the network is much faster than running a conventional field solver. The method is realized by combining two elements: first, using a neural network designed based on PointNet to capture the geometry features of arbitrary interconnect structures; second, minimizing the residual of the governing electrostatic equations with voltage boundary conditions for typical electrostatic analysis so that no computationally expensive simulation data is needed, which is the main highlight of the method. A residual network based architecture is also proposed to improve the method precision. And as we use the point cloud to describe the interconnect structures, the geometry features can be precisely captured and the method is flexible and easily adapted to various interconnect shapes. Numerical results demonstrate that the total error of capacitance produced with Isa-PIPN achieves an average error rate of 5.78% for five highly variant interconnect geometry patterns. Besides the inference time of Isa-PIPN is more than 1000X faster than a 2-D BEM field solver, which is promising for large-scale interconnect capacitance extraction applications.