AnaIR: A Semi-Analytical Green’s Function Inspired Neural Network for Static IR Drop Prediction

IR drop estimation is a classical problem in electronic design and automation (EDA), and is an integral part of the design signoff process. Even after decades of research on efficient circuit simulation and power grid modeling, IR drop estimation tools are still slow and often take hours to compute the IR drop profile even for moderate-sized designs. Given that it is hard to parallelize them, and often there is a need to run them iteratively thousands of times for design space exploration, there is a need for a much faster method, albeit with a minor reduction in accuracy. MLbased approaches have proven to be quite effective in this space; they are at least $\mathbf{1 0 - 1 0 0 X}$ faster. Sadly, they are quite inaccurate as of 2024. Given that they are oblivious to the underlying physics, the errors in predicting hotspots are undesirably high. In this paper, we propose a semi-analytical model AnaIR, which uses Green’s functions to quickly arrive at an approximate IR-drop profile. It is further refined using a small U-Net based neural network to arrive at a much more precise estimate. It is $3.9 \times$ faster than competing work and is $32.6 \%$ more accurate.