Surrogate Model Training

A Surrogate Model Training study step is intended for training a deep neural network and creating a surrogate model that can replace a full finite-element model.

Syntax

model.study(stdname).create(fname, "SurrogateModelTraining");

model.study(stdname).feature(fname).set(pname,value);

Description

Study step.

The following properties are available:

Table 6-166: SurrogateModelProperties Properties
Property	Value	Default	Description
accumtable	new \| name of existing table	new	Output table to use for accumulated probe table.
accumtableall	boolean	false	Use all probes.
adpevals	nonnegative integer	10,000	Surrogate evaluations for optimization, when gpadpoptmethod is set to montecarlo.
computeaction	recompute \| append	recompute	Compute and build surrogate model (recompute), or improve and build it (append).
convinfo	minimal \| normal \| detailed	normal	Surrogate model training log level.
covfunction	se \| matern32 \| matern52 \| nn	matern32	Type of covariance function to use when surrogatemodel is set to gp. Use se for Squared exponential, matern32 for Matérn 3/2, matern52 for Matérn 5/2 and nn for Single-layer neural network.
defsolvergen	global \| all	global	Use global parameters or each parameter tuple for the default solver sequence generation.
errorhandling	immediate \| later	immediate	Stop immediately if error or skip problematic parameters.
globaldnnfunction	new \| name of existing function		Deep Neural Network function to use, when surrogatemodel is set to gp.
globalgpfunction	new \| name of existing function		Gaussian Process function to use, when surrogatemodel is set to gp.
globalpcefunction	new \| name of existing function		Polynomial Chaos Expansion (PCE) function to use, when surrogatemodel is set to gp.
gpadpoptmethod	direct \| montecarlo		Optimization method for error estimation, when surrogatemodel is set to gp.
keepsol	last \| all	last	Keep only last or all model evaluations in memory.
maxgpevals	nonnegative integer	10000	Maximum surrogate evaluations for optimization, when gpadpoptmethod is set to direct.
maxgpiters	nonnegative integer	500	Maximum number of optimization iterations, when gpadpoptmethod is set to direct.
meanfunction	const \| linear \| quadratic	const	Defines the overall trend of the trained function(s) when surrogatemodel is set to gp.
nsolvenonadp	positive integer	20	Number of input points.
objgroup	none or name of table group		The output table group, when surrogatemodel is set to none.
pcesettings	auto \| manual	auto	PCE settings method.
pdistr	boolean	false	Distribute model evaluation.
polydegreespce	positive integer	30	Maximum polynomial degree, when pcesettings is set to manual.
qoisolution	auto \| first \| last \| sum \| min \| max	auto	The solution to use.
qnorm	double (0–1)	0.5	Q norm, when pcesettings is set to manual.
randseed	double	1	Initial random seed, is useseed is set to manual.
reusesol	boolean	false	Reuse solution from previous step.
surrogatemodel	none \| gp \| pce \| dnn	none	Use no (DOE) surrogate model (none) or a Gaussian Process (gp), Polynomial Chaos Expansion (pce), or Deep Neural Network (dnn) surrogate model.
useaccumtable	boolean	false	Use accumulated probe table.
useseed	automatic \| manual \| currenttime	automatic	Random seed time for input parameters sampling.