Hybrid Simulated Annealing and Nelder-Mead Algorithm for Solving Large-Scale ...
B-G-3
1. Group Members
Aakash Rajper 1381-FET/BSEE/F-10
Abdul Haseeb 1382-FET/BSEE/F-10
Badar Jahangee 1383-FET/BSEE/F-10
Danish Javed 1384-FET/BSEE/F-10
Hamza Arshad 1382-FET/BSEE/F-10
2. Used for modeling devices.
Devices are modeled in ATLAS by a set of one to six
PDEs.
These PDEs are converted to approx. non-linear then
linear algebraic equations.
These equations are solved iteratively until a solution
nearest to real values are achieved.
Values for unknowns are calculated at mesh points.
Different solution procedures exhibit different
behavior with respect to convergence, accuracy,
efficiency, and robustness.
3. The two main aspects of convergence are whether a
solution is obtained and how rapidly it is approached.
Accuracy is how closely the computed solution
approximates the true solution.
Efficiency is the time required to produce a solution.
Robustness is the ability of a technique or method to
cope with errors during execution.
Different methods can work better for different
problems. Atlas has captured practical experience in
the form of default methods and parameters that work
well in almost all circumstances.
4. Numerical Methods are given in the METHOD statements of the
input file.
Example “METHOD GUMMEL NEWTON”
There are three types of solution techniques.
1. Decoupled (GUMMEL)
Gummel method will solve for each unknown in turn keeping
the other variables constant, repeating the process until a
stable solution is achieved.
2. Fully coupled (NEWTON)
Newton method solve the total system of unknowns together.
3. BLOCK
Block method will solve some equations fully-coupled while
others are decoupled.
5. A coupled system is formed by two differential
equations with two dependent variables and one
independent variable.
For example:
𝑑𝑦(𝑡)
𝑑𝑥(𝑡)
= 𝑎𝑥(𝑡) + 𝑏𝑦(𝑡)
𝑑𝑥(𝑡)
𝑑𝑡
= 𝑐𝑥(𝑡) + 𝑑𝑦(𝑡)
a, b, c and d are constants while x and y are functions of t
6. In decoupled system one variable is solved while other
are kept constant.
For example
𝜕𝑦(𝑡)
𝜕𝑡
= 𝑎𝑦 𝑡 + 𝑏𝑥(𝑡)
𝜕𝑥(𝑡)
𝜕𝑡
= 𝑐𝑦 𝑡 + 𝑑𝑥(𝑡)
a, b, c and d are constants while x and y are functions of t
7. Gummel Method
Weakly coupled system of equations
Linear convergence
Provide better initial guess
Newton Method
Strongly coupled system of equations
Quadratic convergence
May spend extra time
Requires more accurate initial guess
8. Block Method
Provide faster simulation time
Useful to start with few Gummel iterations
Then switch to Newton to complete the solution
Generates better guess
We compare the performance of algorithm by their
rate of convergence.
9. This model requires the solution of three equations for
Potential
Electron Concentration
Hole concentration
For almost all cases NEWTON method is preferred
and it is the default.
Block method in this model is robust but more time
consuming.
Block method is highly recommended for all
simulations with floating regions (e.g., SOI
transistors)
10. Extra equation is added when Latice heating model is
added to drift diffusion
If we apply Block Method in this level it’ll solve three
eq. by Newton and fourth with Gummel.
Newton Method solves all four equations in a coupled
manner.
Newton is preferred for high temp. while Block is used
for low temp. gradients
11. It requires solution of up to five coupled equations.
Newton and Gummel have same meaning for this
model
Block performs coupled solutions for all five equations
We can switch from Block to Newton for robust
performance.
The switching point is pre-determined.
It can be changed in the METHOD statement.
12. Requires six equations system.
Gummel and Newton solve the equations iteratively.
Block functions initially performs the same as energy
balance and performs lattice heating equation in
decoupled manner.
13. Old Syntax New Syntax
Symbolic newton Method newton
Symbolic gummel Method gummel
14. All numeric settings chosen on METHOD statement
All structure/parameter specification must be before this statement.
All solution specification must be after it.
Fully Coupled Method solves for potential and carriers coupled
(METHOD NEWTON)
Recommended for all cases even including SOI simulations.
De-Coupled method solves potential and carriers sequentially
(METHOD GUMMEL)
Faster for low current cases.
Combined method (METHOD GUMMEL NEWTON)
Runs initial decoupled iterations and switches to coupled.
GUM.INIT parameter controls the number of initial decoupled
iterations.
most robust (but slowest) method.
15. Atlas can solve both electron and hole continuity
equation.
We can make this choice by using CARRIER
parameters.
METHOD CARRIER = 2
Specifies, when a solution for both carriers is required.
This method is default.
METHOD CARRIER = 1 HOLE
For one carrier either electron or hole.
METHOD CARRIER = 0
For potential only.
16. The following cases require ‘METHOD NEWTON
CARRIER = 2’ to be set for isothermal drift diffusion
simulations
Current boundary conditions
Distributed or Lumped external elements
AC analysis
Impact ionization
Both Block or Newton are permitted for lattice heat
and energy balance.