Scalable Compact Models for Complex High-Speed Systems

Scalable Compact Models for

Complex High-Speed SystemsComplex High-Speed Systems

Francesco Ferranti, Tom Dhaene

Department of Information Technology (INTEC),

Ghent University – IBBT

WMC: Advanced techniques for electromagnetic-based model generation

Introduction

Scalable Macromodels

OUTLINE

WMC: Advanced techniques for electromagnetic-based model generation IMS2012, Montreal, June 17-22, 2012 2

Numerical examples• Data-driven PMOR

• Model-driven PMOR

Conclusions

Introduction


OUTLINE




Conclusions

Introduction

input output

fluid dynamicselectronics

telecom


input outputout = f(in)

automotivechemistry biomodeling geology

Introduction

AerodynamicsAutomotive Electronics MetallurgyChemistry

Designvariables

width, temperature,

angle, frequency, ...

Responsevariables

lift, S-parameters,

pressure, stress, ...

Simulation Model

Fluent®, HSPICE®, CST®,

Comsol®, Abaqus®, ...

Costlyp2

Adaptive Modeling


OptimizationSensitivity

AnalysisPrototyping

CAD/CAM/CAE

Environment

Designvariables

Responsevariables

Cheap

Surrogate Model / Metamodel

Neural network, Kriging, SVM, rational function, spline,...

p3

p1

p3

p1

p2

Configurable infrastructure

Distributed Computing

Introduction


Designvariables

width, temperature,


Responsevariables

lift, S-parameters,


Simulation Model



Costlyp2

Adaptive Modeling



AnalysisPrototyping

CAD/CAM/CAE

Environment

Designvariables

Responsevariables

Cheap



p3

p1

p3

p1

p2



Introduction


Introduction


Introduction


Design process

� several decisions

� materials

� geometrical dimensions

shape

Introduction


� shape

� constraints

� space

� cost

� performance

Simulators � implementation of models

� describe systems behavior

� help designers

Measurements

� post tuning

� verification

� help designers

Design process


Introduction


OUTLINE




Conclusions

A typical design process requires

� Multiple simulations (measurements)

� different design parameters values (e.g. layout features)

� design space optimization

� design space exploration

Scalable macromodels


� design space exploration

� sensitivity analysis

A typical design process requires


� computationally expensive (time and memory)



� Can we do better?

� Yes

� By scalable macromodels

discretisation by

simulators

modeling

model

PDE


real world


Model-driven PMOR

sX= AX+BUY=CX+DU

U(s,g) Y(s,g)

H(s,g) U(s,g) Y(s,g)

scalable macromodel

PDE


discretisation by

simulators

modeling

model

PDE

real world


Data-driven PMOR

data

simulations

Model-driven PMOR

sX= AX+BUY=CX+DU

U(s,g) Y(s,g)

PDE


scalable macromodel


measurements

modeling

model

PDE

real world

discretisation by

simulators


Data-driven PMOR

data

Model-driven PMOR

sX= AX+BUY=CX+DU

U(s,g) Y(s,g)


scalable macromodel

PDE

PMOR concepts

Two design space grids are used in the modeling process

� estimation grid

� validation grid Design space



N

n

ngg 1

)()(

==

discretisation by

simulators

modeling

model

PDE


real world


Model-driven PMOR

sX= AX+BUY=CX+DU

U(s,g) Y(s,g)


scalable macromodel

PDE

PRIMA

LAGUERRE SVD

FAST TBR

HYBRID TBRModel Order Reduction

Interpolation

Least Squares

SYSTEM MATRICES

TRANSFER FUNCTIONS

Model-driven PMOR


A B

C D

HAr Br

Cr Dr

Hr

EIGENSPACE PROJECTION

MOR

PROJECTION MATRICES

PRIMA

LAGUERRE SVD

FAST TBR


Interpolation

Least Squares

SYSTEM MATRICES

TRANSFER FUNCTIONS

Model-driven PMOR



PROJECTION MATRICES

A B

C D

HAr Br

Cr Dr

Hr

A B

C D

HAr Br

Cr Dr

Hr

A B

C D

HAr Br

Cr Dr

Hr

A B

C D

HAr Br

Cr Dr

Hr

PRIMA

LAGUERRE SVD

FAST TBR


SYSTEM MATRICES

TRANSFER FUNCTIONS

Model-driven PMOR

Interpolation

Least Squares



PROJECTION MATRICES

model-driven PMOR

accuracy

efficiencyHr(s,g) stability guaranteed

passivity guaranteed

A B

C D

HA B

C D

HA B

C D

HA B

C D

H

Data-driven PMOR

discretisation by

simulators

modeling

model

PDE

real world

measurements


Data-driven PMOR

data

simulations

Model-driven PMOR

sX= AX+BUY=CX+DU

U(s,g) Y(s,g)

PDE


scalable macromodel

VECTFIT

MULTIPORT

PASSIVITY

STABILITY

Rational Fitting

Methods

SYSTEM MATRICES

POLES/RESIDUES

TRANSFER FUNCTIONS

Data-driven PMOR

Interpolation

Least Squares


data-driven PMOR

Hr(s,g)

NOISY DATA

accuracy

efficiency

stability guaranteed


scattered data

POLES/RESIDUES

Data-driven PMOR

VECTFIT

MULTIPORT

PASSIVITY

STABILITY

Rational Fitting

Methods

SYSTEM MATRICES

POLES/RESIDUES

TRANSFER FUNCTIONS

Interpolation

Least Squares


Hr(s,g))()())()((),(

1

r gDgBgAIgCgH +−=−

ss

∑=

−

=

P

p pass

1

r)(

)(),(

g

gQgH

scattered data

data-driven PMOR

NOISY DATAPOLES/RESIDUES

Introduction


OUTLINE




Conclusions

Introduction


OUTLINE




Conclusions

Data-driven PMOR

VECTFIT

MULTIPORT

PASSIVITY

STABILITY

Rational Fitting

Methods

SYSTEM MATRICES

POLES/RESIDUES

TRANSFER FUNCTIONS

Interpolation

Least Squares


Hr(s,g))()())()((),(

1

r gDgBgAIgCgH +−=−

ss

∑=

−

=

P

p pass

1

r)(

)(),(

g

gQgH

scattered data

data-driven PMOR

NOISY DATAPOLES/RESIDUES

3D example: Interconnection structure

� Mobile phone application

� 10 lines (20 ports)

� etched ground plane

Data-driven PMOR example


� CPU time model estimation

� CPU time model validation Design space



Poles



Output



Output



� Design optimization

� minimum S, minimum α

� |S13(s,S,α)| <= -22 dB over [0 - 20] GHz





� |S13(s,S,α)| <= -22 dB over [0 - 20] GHz

� initial values





� |S13(s,S,α)| <= -22 dB over [0 - 20] GHz

� initial values

� optimal values



� optimal values




� |S13(s,S,α)| <= -22 dB over [0 - 20] GHz

� initial values

� optimal values


� optimal values

Introduction


OUTLINE




Conclusions

PRIMA

LAGUERRE SVD

FAST TBR


SYSTEM MATRICES

TRANSFER FUNCTIONS

Model-driven PMOR

Interpolation

Least Squares



PROJECTION MATRICES

model-driven PMOR

accuracy

efficiencyHr(s,g) stability guaranteed


A B

C D

HA B

C D

HA B

C D

HA B

C D

H

3D example: Spiral inductor

Model-driven PMOR example


Original order = 801

Reduced order = 91



Ly= 0.46 mm Ly= 0.93 mm




Lx= 0.63 mm


Introduction


OUTLINE




Conclusions

Conclusions

input output

fluid dynamicselectronics

telecom


input outputout = f(in)

automotivechemistry biomodeling geology

Conclusions


Designvariables

width, temperature,


Responsevariables

lift, S-parameters,


Simulation Model



Costlyp2

Adaptive Modeling



AnalysisPrototyping

CAD/CAM/CAE

Environment

Designvariables

Responsevariables

Cheap



p3

p1

p3

p1

p2



Conclusions


Designvariables

width, temperature,


Responsevariables

lift, S-parameters,


Simulation Model



Costlyp2

Adaptive Modeling



AnalysisPrototyping

CAD/CAM/CAE

Environment

Designvariables

Responsevariables

Cheap



p3

p1

p3

p1

p2




Multiple design variables

Compact models

Conclusions


scalable macromodel


Compact models

Efficient design activities (excellent speed-ups)


� Design space optimization, exploration, sensitivity analysis


Time-domain simulations

• Non-linear drivers and receivers

Stochastic modeling

Conclusions


scalable macromodel


Stochastic modeling

• impact of manufacturing tolerances

Models from measurements

• noise to handle

Applications in different domains

Questions


Documents

Scalable Compact Models for Complex High-Speed Systems