Channel Characteristic-Based Deep Neural Network Models for Accurate Eye Diagram Estimation in High Bandwidth Memory (HBM) Silicon Interposer

 D. Lho et al., "Channel Characteristic-Based Deep Neural Network Models for Accurate Eye Diagram Estimation in High Bandwidth Memory (HBM) Silicon Interposer," in IEEE Transactions on Electromagnetic Compatibility, vol. 64, no. 1, pp. 196-208, Feb. 2022, doi: 10.1109/TEMC.2021.3081713.

• Deep Neural Network (DNN) for eye height/width estimation of high bandwidth memory (HBM) silicon interposer channels.
• The abstract only talks about how they method has lower error rates than other models but they don't specify the actual uncertainty bounds.  I wish they would say, our method was able to estimate the eye width by +/- x ps or whatever.
• due to density of routing, crosstalk is a big issue.
• The channel used was just a stripline or microstrip.  
• After I read about the KBNN (knowledge based neural net) I can't stop seeing it as the best idea I've seen so far.

Knowledge-Based Neural Networks for Fast Design Space Exploration of Hybrid Copper-Graphene On-Chip Interconnect Networks

 R. Kumar et al., "Knowledge-Based Neural Networks for Fast Design Space Exploration of Hybrid Copper-Graphene On-Chip Interconnect Networks," in IEEE Transactions on Electromagnetic Compatibility, vol. 64, no. 1, pp. 182-195, Feb. 2022, doi: 10.1109/TEMC.2021.3091714.

• This paper does Artificial neural network (ANN) of a per-unit-length (p.u.l.) parameters of a copper-graphene on-chip interconnects.  
• I did not know that these were difficult to calculate through a full wave simulation.  
• The introduction of the paper makes a good case that the TEM pul simulations are very time intensive.
• PCB transmission lines have had lots of attention and there are lots of approximations as to their behavior. Maybe using a ANN here is a good idea since cooper-graphene interconnects are not as well understood, so creating such a model is quite valuable.
• It would be interesting to do the same work with regular PCB transmission lines and see how the model compares to the common TL equations. This would be a good project for a student to do.
• Knowledge based neural networks (KBNN)
• The objective is to predict the R,L,C matrices for copper graphene transmission lines
• Source difference KBNN
• 1) Come up with some empirical equations for calculating RLC matrices from the input parameters.  There are some ok models out there but they are not perfect.
• 2) Using an EM solver, cover the input parameter space with lots of simulations
• 3) Find the error in the RLC predictions between the EM solver and the empirical equations.
• 4) Fit a ANN to the error will be much more easy since the empirical equation models already captured much of the known knowledge of the system under study.
• 5) The final RLC model is composed of the empirical equations corrected by the ANN fit of the error.
• This marries the best of the empirical equations with the strength of the ANN.  Since the ANN fit of the error is much simpler, it requires fewer data points to achieve a given level of accuracy.
• 
• This is a great idea that I am very impressed with.
• Used the MATLAB machine learning toolbox 19a
• Over this is really the first useful application of machine learning that I have seen for applications in Signal Integrity.

3/16/23 Edit: Found this: https://www.mathworks.com/matlabcentral/fileexchange/93675-knowledge-based-neural-networks