M. Bellini, ABB Corporate Research, MOS-AK workshop, ESSDERC, Lausanne 2016

Use and extraction of compact models for EMI / EMC simulations of power devices

Slide 2 © ABB

•  Motivation •  Physics of Power Diodes •  Review of Existing Models •  Improved Model Power Diode Model •  IGBT Model and Parameter Extraction •  Results and Discussion

Motivation

Slide 3 © ABB

•  Accurate and fast compact model needed for predictive EMI/EMC simulations. No clear standard power diode models. Parameter extraction is very complex, because of coupled phenomena.

•  The proposed diode model can accurately simulate reverse recovery in 120 ms, in contrast with 5-10 s for FEM SPICE models and 2-10 minutes for TCAD.

•  The SCR expansion and the excess charge simulated by the circuit model agree closely to calibrated TCAD simulations. But the high precision in reverse recovery is obtained at the expense of DC characteristics (less important for EMI/EMC).

Requirements & Motivation

Slide 4 © ABB

Typically used by designers: •  PSPICE •  Simmetrix (Verilog-A)

Being evaluated at CRC: •  Xyce •  Qucs

Requirements: •  GUI, device library (Verilog A) •  Command line operation (model fitting) •  Model / Circuit convergence debugging capability

Typical simulation tools

Slide 5 © ABB

Power Diodes: Characteristics

Device under test:

VRRM 4500 [V] IFAV 1650 [A] Electron irradiation

Energy

Dose

1-5

5-20

[MeV]

[kGy]

He irradiation

Energy

Dose

5-12

1010-1011

[MeV]

[cm-2]

Slide 6 © ABB

Power Diodes: Structure and Physics

The depth of the semiconductor is used to sustain the applied voltage. The maximum doping is limited by the cosmic ray failure mechanism.

anode

cathode

The bulk region of the semiconductor is chosen with the appropriate n- doping, which is generally in the 1013 cm-3 range The p+ and n+ contact are diffused at the top and bottom of the wafer.

Slide 7 © ABB

Power Diodes: Structure and Physics

Modulation reduces the on-state voltage drop to a few volts for currents of thousand Ampere. However, excessive charge storage in the drift region leads to very high switching losses.

2

2

dxpdDp

dtdp

+−=τ Ambipolar Diffusion Equation (ADE)

Slide 8 © ABB

Power Diodes: Structure and Physics

Switching circuit used for measurements

Slide 9 © ABB

Power Diodes: Structure and Physics

Switching circuit used for measurements

Slide 10 © ABB

Power Diodes: Structure and Physics

Switching circuit used for measurements

Slide 11 © ABB

Power Diodes: Structure and Physics

Switching circuit used for measurements

Slide 12 © ABB

Power Diodes: Structure and Physics

Switching circuit used for measurements Switching circuit used for measurements

Slide 13 © ABB

Power Diodes: Structure and Physics

Switching circuit used for measurements

Slide 14 © ABB

Power Diodes: Structure and Physics

Switching circuit used for measurements

Slide 15 © ABB

•  e.g. FLECS

•  Lookup table models

•  can only describe voltage drops and commutation losses

•  no transient behavior

•  very widely used for system-level simulation

Modelling Approaches Functional Models

October 6, 2016

Slide 16 © ABB

•  e.g. Hefner, Kraus

•  Physical

•  Few parameters (doping, thickness, lifetime, transit time)

•  Limited to moderate accuracy àcomplexity of the approximation and implementation

•  Accurate models have typically convergence or speed problems

Modelling Approaches Approximate Solution Models (Physical Models)

October 6, 2016

Approximate solution of the Ambipolar Diffusion Equation (ADE)

2

2

dxpdDp

dtdp

+−=τ

Slide 17 © ABB

•  e.g. Strollo, Bryant

•  Physical behavior

•  Better accuracy but implementation in circuit simulators (if implemented with RC networks it can cause convergence degradation)

•  Possible oscillation / series truncation problems

Modelling Approaches Laplace Transform Models

October 6, 2016

( ) ( )ttxpp

xtxpD

∂

∂+=

∂

∂ ,,2

2

τ

( ) ( ) ( )( ) ⎟

⎟⎠

⎞⎜⎜⎝

⎛

−

−=∑

∞

=12

10

cos,xxxxktptxp

k kπ

x1 and x2 are the boundaries of the modulated region.

Slide 18 © ABB

•  e.g. Ma, Lauritzen

•  Physical behavior, potential for good accuracy

•  Very high speed: it’s possible to simulate large circuits

•  Very good convergence properties

•  Parameter extraction can be quite challenging

Modelling Approaches Lumped Charge Models

October 6, 2016

Slide 19 © ABB

Modelling Approaches Numerical Solution Models (FE, FD)

October 6, 2016

•  e.g. Buiatti

•  Potential for very high accuracy

•  Very low speed: Only small circuits can be simulated

•  Circuit simulators are not the best tools for FD or FE models: use of a TCAD simulator can improve precision / speed

•  Parameter extraction can be extremely challenging

Slide 20 © ABB

Lumped Charge Model Ma-Lauritzen Model Derivation

October 6, 2016

Slide 21 © ABB

Avalanche generation is introduced in the model. The simplification are balanced by a field dependent recombination rate.

Lumped Charge Model Extended Ma-Lauritzen Model Derivation

October 6, 2016

Slide 22 © ABB

The multi-objective genetic algorithm NSGA-II is used to identify the optimal tradeoff between DC and RR (Pareto Frontier).

Parameter Extraction Procedure Multi-objective genetic algorithm to automate parameter extraction

October 6, 2016

8 parameters, 2 objectives

Slide 23 © ABB

Excellent accuracy in RR (good for EMI simulations). DC shows excessively resistive behavior.

Results: DC and Reverse Recovery Comparison with measured electrical characteristics

October 6, 2016

Slide 24 © ABB

The width of the SCR and the excess charge closely match calibrated TCAD.

Results: Internal Parameters Match with internal quantities calculated by TCAD

October 6, 2016

Slide 25 © ABB

Parameter Extraction/Fitting Procedure Multi-objective genetic algorithm to automate parameter extraction

October 6, 2016

Optim

izer - Measurement - Simulation

The optimizer varies the model parameters’ until the F.O.M. confirms that simulation matches data.

Slide 26 © ABB

IGBT-DIE 5SMY 12K1721, Ic = 100 A, Multiobjective optimization: 5 targets, 14 parameters

Parameter Extraction Procedure Multi-objective genetic algorithm to automate parameter extraction

October 6, 2016

Gat

e ch

arge

Tu

rn-o

n Ic

O

n-st

ate

trans

fer

Turn

off

Vce

The model scales well with the load current

Good agreement with the remaining characteristics

Slide 27 © ABB

Line impedance stabilization network can be simulated in half time and without convergence issues on a circuit with 70’000 variables.

Results: System EMI/EMC simulations Low Voltage AC Drive Noise Simulations incl. cable models

October 6, 2016

Slide 28 © ABB

•  Wide variety of users / applications •  Both GUI and command line operation desired •  Verilog A integration necessary to model a variety of components •  Debugging capability could dramatically shorten model development

time

Conclusions

October 6, 2016