TY - GEN
T1 - Impedance Response Extrapolation of Power Delivery Networks using Recurrent Neural Networks
AU - Bhatti, Osama Waqar
AU - Swaminathan, Madhavan
N1 - Funding Information:
response to yield an interesting comparison. Horizontal line in the Fig. 5 shows the cut-off point on wcj axis. We observe that the predicted poles occur at locations spatially near and sometimes exact to the poles fitted from the actual response. V. CONCLUSION We show that frequency response samples are correlated in frequency space. We treat the frequency response of a power delivery network as sequenced data and employ recurrent neural networks to extrapolate the impedance response over frequency. Our method provides 66% bandwidth extension within an MSE error of 8x10−3. Future work includes developing methods for further extrapolation and defining a confidence bound for the error. ACKNOWLEDGMENT This material is based upon work supported by the National Science Foundation under Grant No. CNS 16-24810 - Center for Advanced Electronics through Machine Learning (CAEML).
Funding Information:
This material is based upon work supported by the National Science Foundation under Grant No. CNS 16-24810 - Center for Advanced Electronics through Machine Learning (CAEML)
Publisher Copyright:
© 2019 IEEE.
PY - 2019/10
Y1 - 2019/10
N2 - Often times the impedance response of a power delivery network needs to be extrapolated to determine if any resonances occur in the vicinity or outside of the band-limited response provided. If the circuit models are unavailable, this can become a cumbersome exercise. We propose a machine learning aided method using long short term memory recurrent neural networks to extrapolate the response in frequency thereby avoiding extensive simulations and saving computational time as opposed to EM solvers. Results show that the accuracy in the prediction is good with a mean square error of 0.008.
AB - Often times the impedance response of a power delivery network needs to be extrapolated to determine if any resonances occur in the vicinity or outside of the band-limited response provided. If the circuit models are unavailable, this can become a cumbersome exercise. We propose a machine learning aided method using long short term memory recurrent neural networks to extrapolate the response in frequency thereby avoiding extensive simulations and saving computational time as opposed to EM solvers. Results show that the accuracy in the prediction is good with a mean square error of 0.008.
UR - http://www.scopus.com/inward/record.url?scp=85084598690&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85084598690&partnerID=8YFLogxK
U2 - 10.1109/EPEPS47316.2019.193198
DO - 10.1109/EPEPS47316.2019.193198
M3 - Conference contribution
AN - SCOPUS:85084598690
T3 - 2019 IEEE 28th Conference on Electrical Performance of Electronic Packaging and Systems, EPEPS 2019
BT - 2019 IEEE 28th Conference on Electrical Performance of Electronic Packaging and Systems, EPEPS 2019
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 28th IEEE Conference on Electrical Performance of Electronic Packaging and Systems, EPEPS 2019
Y2 - 6 October 2019 through 9 October 2019
ER -