Model Reduction
Use a Model Reduction study node () to create a reduced-order model (ROM) based on a time-dependent or frequency-domain simulation (see Reduced-Order Modeling). Reduced-order models are usually thought of as computationally inexpensive mathematical representations that provide the ability to run faster simulations using a small model that captures the behavior of the original model. Constructing the reduced-order models can be computationally expensive as it requires accumulating a large number of responses to input excitations (modes). The creation of a reduced-order model can typically be divided into two steps: production of training data and model building. The resulting model can then be used for repeated simulations. The output from the training study at the specified study step is used as the source for training data (for the modal solver this corresponds to eigenvectors).
To add a Model Reduction node, first select Reduced-Order Modeling in the Show More Options dialog box.
There can only be one Model Reduction node in a study. When you copy a Model Reduction node, it is possible to paste it into another study without a Model Reduction node.
The following steps are the main steps needed to set up a model reduction study:
1
2
3
4
Select to create or update an instance of the reduced model under Global Definitions>Reduced Models for online use.
5
6
7
Click the Compute button () (or press F8) at the top of the Settings window to produce an instance of the reduced model under Global Definitions>Reduced Models. When an instance has been created, you can also click the Update Solution button ().
The Settings window contains the following sections:
Model Reduction Settings
In this section, you specify how to run the model reduction.
Reduced-Order Method
From the Method list, choose one of the following model-reduction methods:
The Modal method (the default) supports inputs and outputs and makes it possible to export ROM matrices.
The AWE (asymptotic waveform evaluation) method is an alternative reduced-order model that provides a fast-frequency sweep (an advanced interpolation) and supports outputs only.
Training Study (Modal Method Only)
From the Training study list, choose an existing study for the basis functions (training data) or choose None. The study must contain an eigenvalue solver.
From the Defined by study step list, choose Automatic (the default) to use the last applicable study step in the study, or select any of the applicable study steps (such as Eigenvalue).
From the Compute list, choose Initially (the default) to use the initially computed eigenvalue solution, or choose Always to always compute the eigenvalue solution. The default setting is computationally efficient but may not be valid if you have made changes to the model that affect the eigenvalue solution.
Unreduced Model Study
From the Unreduced model study list, choose the study that solves the unreduced model (for example, a time-dependent study). By default, the last enabled study step or study reference that is compatible with the selected method, if any, is used as the definition of the unreduced model.
From the Defined by study step list, select any of the applicable study steps (such as Time Dependent).
From the Reduced-order model list, choose New to create a new reduced-order model, or choose any existing and compatible reduced-order model (available under Global Definitions>Reduced-Order Modeling).
If the model-reduction method is Modal, select the Ensure reconstruction capability check box to enable reconstruction of the unreduced solution vector (for the AWE method, reconstruction is always enabled). The reduced-order model can then also assign reconstructed values to some of the DOFs not solved for. This is controlled by the table with Reconstruction and Reduced-order model columns in the Physics and Variables Selection section in the destination study. There is a row for each physics interface that is not solved for and that has one or several corresponding reduced-order models. The Reconstruction column shows the physics interface name. The list in the Reduced-order model column determines which reduced-order (if any) should reconstruct the fields for this physics interface.
For the Modal method, select the Store reduced matrices check box to store the reduced matrices from the model reduction in the solution data for exporting state-space matrices, for example.
For the AWE method, enter a value for the Relative tolerance for adaptation (default: 0.01).
Model Control Inputs
In this section, you define the model control inputs. The Model Control Inputs table consists of three columns: Reduced model input, Use, and Training expression. The Reduced model input column shows all the variables defined in the Global Reduced Model Inputs node under Global Definitions. When the variable is added to the Global Reduced Model Inputs it is automatically added to the Model Control Inputs table. The Use column controls which of the defined variables that should be used. In the Training expression column, enter a training expression that is compatible with the training study step.
Outputs
In the Outputs section, add outputs for the reduced-order model. You can add output variables by clicking the Add Expression () and Replace Expression () buttons to search through a list of predefined expressions. If you do not add a name in the Variable column, the output is assigned a default variable name in the Reduced Model node. For the AWE method, you can also add a scaling factor for each output in the Scale column and select the check box (selected by default) in the Use for adaptation column to include it in a mesh adaptation.
Results While Solving
This section is available for model reduction in the frequency domain only (AWE selected from the Method list).
See Results While Solving in the Common Study Step Settings section. Also see Getting Results While Solving.
Thermal Controller, Reduced Order Model: Application Library path COMSOL_Multiphysics/Multiphysics/thermal_controller_rom.