eldo_user

Monte Carlo Analysis
Model-Based Monte Carlo Simulation
requirements. SSD indicates a different approach compared to the standard Monte Carlo or
Latin Hypercube methods, where the specific sampling plan does not require too many
simulations. The SSD approach has to be understood now as an ensemble of modeling
techniques designed to accelerate the variability analysis.
In model-based Monte Carlo settings, the variability analysis is composed of three distinct
steps:
• The first one is the sampling phase where a space-filling design is used to collect
variations of the output responses of a circuit from the input variations.
• The modeling phase is formulated as a pure regression (or interpolation) problem.
• And the final variability analysis is directly performed on the surrogate function.
The key points of the SSD approach are as follows:
• You provide the number of runs that you want Eldo to perform, NRUN_MAX, based on
how much CPU-time you want to spend, and Eldo determines a sampling plan with
some space filling properties.
• You may also specify the number of test runs that you want Eldo to perform, using the
.MC parameter SSD_NRUN_TEST. The test runs are used to compute accuracy
metrics, contained in the .mcm file, which you can use to determine the quality of the
surrogate model. Refer to “Model Adequacy Checking for Model-Based MC” on
page 476 for more information.
Note
By default, Eldo does not perform any test runs. Prior to AMS release 15.3, Eldo
automatically determined how many test runs to perform based on the value of
NRUN_MAX.
• The complexity of the surrogate model for the SSD sampling method
(SAMPLING=SSD) is controlled with the .MC parameter SSD_COMPLEXITY,
defined as multiple choice from the simplest model (0) to the more complex (1) which is
the default.
We assume that each circuit response y i , i=1, …, N is simulated with d variables x 1 , …,
x d given the realizations {y i , x 1i , …, x di }. The goal is to estimate a function s that
satisfies a regression model:
over some domain: containing the input data. The additive stochastic term ε has
its expected value defined to be zero, and defines the dependence of the output variable
on quantities other than the selected input variables.
460
Eldo® User's Manual, 15.3