Generating Kriging Models with PySMO#

Note

The IDAES surrogate API is a wrapper around the original PySMO surrogate interface.

pysmo_surrogate.PysmoKrigingTrainer trains ordinary Kriging models. Interpolating kriging models assume that the outputs $$\hat{y}\in\mathbb{R}^{m\times1}$$ are correlated and may be treated as a normally distributed stochastic process. For a set of input measurements $$X=\left\{ x_{1},x_{2},\ldots,x_{m}\right\} ;x_{i}\in\mathbb{R}^{n}$$, the output $$\hat{y}$$ is modeled as the sum of a mean $$\left(\mu\right)$$ and a Gaussian process error,

$$$\hat{y_{k}}=\mu+\epsilon\left(x_{k}\right)\qquad k=1,\ldots,m \qquad\quad$$$

Kriging models assume that the errors in the outputs $$\epsilon$$ are correlated proportionally to the distance between corresponding points,

$$$\text{cor}\left[\epsilon\left(x_{j}\right),\epsilon\left(x_{k}\right)\right]=\exp\left(-\sum_{i=1}^{n}\theta_{i}\mid x_{ij}-x_{ik}\mid^{\tau_{i}}\right)\qquad j,k=1,\ldots,m;\:\tau_{i}\in\left[1,2\right];\:\theta_{i}\geq0\qquad\quad\label{eq:corr-function}$$$

The hyperparameters of the Kriging model $$\left(\mu,\sigma^{2},\theta_{1},\ldots,\theta_{n},\tau_{1},\ldots,\tau_{n}\right)$$ are selected such that the concentrated log likelihood function is maximized.

Basic Usage#

To generate a Kriging model with PySMO, the pysmo_surrogate.PysmoKrigingTrainer trainer is instantiated and initialized with the desired optional arguments, and the training function train_surrogate is called:

# Required imports
>>> from idaes.core.surrogate.pysmo_surrogate import PysmoKrigingTrainer
>>> import pandas as pd

# Load dataset from a csv file

# Define the input and output labels
input_labels = ['X1', 'X2']
output_labels = ['Y1','Y2']

# Create the Kriging trainer object
>>> krg_trainer = PysmoKrigingTrainer(input_labels=input_labels, output_labels=output_labels, training_dataframe = data_training)

# Set desired PySMO kriging options

# Train the model
>>> krg_train = krg_trainer.train_surrogate()


Configuration Options#

PysmoKrigingTrainer takes the following optional arguments:

PySMO Kriging options#

Option

Configuration argument

Description

Whether or not numerical gradients should be used in training. This choice determines the algorithm used to solve the problem.
- True: The problem is solved with BFGS using central differencing with $$\Delta=10^{-6}$$ to evaluate numerical gradients.
- False: The problem is solved with Basinhopping, a stochastic optimization algorithm.

regularization

PysmoKrigingTrainer.config.regularization

Boolean argument which determines whether model regularization is applied during training.
- When regularization is turned on, the resulting model is a regressing kriging model.
- When regularization is turned off, the resulting model is an interpolating kriging model.
Default is True.

Output#

The result of the pysmo_surrogate.PysmoKrigingTrainer method is a python object containing information about the optimal Kriging hyperparameters $$\left(\mu,\sigma^{2},\theta_{1},\ldots,\theta_{n}\right)$$ and different error and quality-of-fit metrics such as the mean-squared-error (MSE) and the $$R^{2}$$ coefficient-of-fit.

Surrogate Visualization#

For visualizing PySMO-trained surrogates via parity and residual plots, see Visualizing Surrogate Model Results.

Building the IDAES Surrogate Object#

To add the model to an IDAES flowsheet or generate model predictions, the SurrogateTrainer object needs to be transformed into an IDAES SurrogateObject object. This is done by calling PySMOSurrogate and passing the generated surrogate expressions, along with variable labels and optionally the bounds:

>>> surr = PysmoSurrogate(krg_train, input_labels, output_labels, input_bounds)


The resulting PysmoSurrogate object may be saved to (and reloaded from) a JSON file; for details, see the PySMO main page.

Prediction with PysmoKrigingTrainer models#

Once the Kriging model has been trained and the SurrogateObject object created, predictions for values at previously unsampled points x_unsampled can be evaluated by calling SurrogateObject’s evaluate_surrogate() function on the unsampled points:

>>> y_unsampled = surr.evaluate_surrogate(x_unsampled)


Flowsheet Integration#

The final Kriging model can be passed into a process flowsheet using the IDAES SurrogateBlock option. The following code snippet demonstrates how a saved Kriging model may be integrated directly into an IDAES flowsheet:

# Required imports
>>> from pyomo.environ import Var, ConcreteModel, Constraint, SolverFactory, Objective, minimize
>>> from idaes.core import FlowsheetBlock
>>> from idaes.core.surrogate.pysmo_surrogate import PysmoSurrogate
>>> from idaes.core.surrogate.surrogate_block import SurrogateBlock

# Create a Pyomo model
>>> m = pyo.ConcreteModel()
>>> m.fs = FlowsheetBlock(default={"dynamic": False})

# create input and output variables
>>> m.fs.X1 = Var(initialize=0, bounds=(0, 5))
>>> m.fs.X2 = Var(initialize=0, bounds=(0, 5))
>>> m.fs.Y1 = Var(initialize=0)
>>> m.fs.Y2 = Var(initialize=0)

# create list of surrogate inputs and outputs for flowsheet
>>> inputs = [m.fs.X1, m.fs.X2]
>>> outputs = [m.fs.Y1, m.fs.Y2]

# create the Pyomo/IDAES block that corresponds to the surrogate
>>> m.fs.surrogate = SurrogateBlock(concrete=True)
>>> surrogates_obj =PysmoSurrogate.load_from_file('krg_surrogate.json') # krg_surrogate.json is an existing surrogate JSON file containing a kriging model
>>> m.fs.surrogate.build_model(surrogates_obj, input_vars=inputs, output_vars=outputs)
>>> m.fs.surrogate.pprint()

# Set the variable Y1 as the model objective
>>> m.fs.obj = Objective(expr=m.fs.Y1, sense=minimize)

# Solve the model
>>> solver = SolverFactory('ipopt')
>>> res = solver.solve(m, tee=True)
>>> m.fs.display()
`

For an example of optimizing a flowsheet containing a PySMO-trained Kriging surrogate model, see the Autothermal reformer flowsheet optimization example.

References:#

[1] Forrester et al.’s book “Engineering Design via Surrogate Modelling: A Practical Guide”, https://onlinelibrary.wiley.com/doi/pdf/10.1002/9780470770801

[2] D. R. Jones, A taxonomy of global optimization methods based on response surfaces, Journal of Global Optimization, https://link.springer.com/article/10.1023%2FA%3A1012771025575