DC25_PAPER_Track14_ImpedanceProfilePredictionandClassification_Hassab_V2.pdf

1、DesignCon 2025Impedance Profile Prediction andClassification for PCBbased PDNDecoupling Using AutoencodersYoucef Hassab,Hamburg University of Technologyyoucef.hassabtuhh.deJan Severin Heling,Hamburg University of Technologyjan.hesslingtuhh.deMorten Schierholz,Hamburg University of Technologymorten.s

2、chierholztuhh.deIhsan Erdin,Celestica Inc.ierdinieee.orgJayaprakash Balachandran,dMatrix Corp.jayapdmatrix.aiChristian Schuster,Hamburg University of Technologyschustertuhh.deAbstractIn this work,a machine learning(ML)based method for fast full impedance profile prediction to bypass orspeed up print

3、ed circuit board(PCB)based power delivery network(PDN)decoupling using autoencodersis proposed.Autoencoders,a type of neural network used to learn efficient representations of data,areshowcased for power integrity applications.In the proposed approach,autoencoders are trained usingavailable datasets

4、 from previous designs to predict the full impedance profile of PDNs.For a new,unseendesign space,the trained models are further tuned using few new samples to make reliable predictionsof the impedance.This approach should lead to a speedup of the decoupling process by allowing fastelectromagnetic(E

5、M)behavior prediction using autoencoders.Moreover,the autoencoders are used tofind similar previous designs to apply previous decoupling strategies to bypass the design process.For thisinvestigation,multiple cases of rectangular and irregularly shaped PCBs with up to 8 metal layers and viaarrays wit

6、h up to 14 ports are simulated up to a frequency of 1 GHz.The used datasets,generated using aphysicsbased(PB)simulation tool,are available on the SI/PIDatabase from the Hamburg University ofTechnology(https:/www.tet.tuhh.de/en/sipidatabase).BiographyYoucef HassabYoucef Hassab was born in Oran,Algeri