Skip to content

Initialization and Training of a Surrogate Model

🎬 Overview

Surrogate models are standalone files in VirtualLab Fusion that can be attached to compatible components (e.g., metalens components) to compute realistic meta-structures. This tutorial describes how to define, train, and use custom surrogate models.

  1. Initialize the surrogate document – load the template or start from a compatible component

  2. Configure optical parameters – define media, wavelength and incidence angle ranges

  3. Configure fabrication parameters – set meta-atom geometry (nanopillars or nanofins) and fix variation parameter space

  4. Train the model – run RCWA calculations and generate the neural network

🚀 Step-by-Step Tutorial

Step 1: Initialize Document

To train a custom surrogate model:

  • Open a new Surrogate Model via File/New/Surrogate Model or load one of the following examples:
  • Link.
  • Alternatively, if you already have a compatible component (e.g., a Metalens [PCA] component), navigate to “Initialize Training” and click “Train Surrogate”.

A new document window will open.

Initialize surrogate model
Initialize surrogate model
Alternative initialization over metalens component.
Alternative initialization over metalens component.
The surrogate model document.
Configure the solution space for the surrogate model

Step 2: Configure Optical System Parameters (Left Panel)

The left panel contains parameters defining the optical system:

  • Media before and after the component
  • Design wavelength
  • Input angles

Parameter Definition:

Parameter Definition Method Notes
Input Angle Range + Steps Centered around zero
Wavelength Range + Steps Centered around the wavelength specified at the top of the document. The Range parameter appears only when Steps > 1.

The surrogate model interpolates within this parameter space. Increasing the Steps parameter improves accuracy at the expense of longer training time.

Step 3: Configure Fabrication Parameters (Right Panel)

The right panel defines meta-cell fabrication parameters. As of Release 2026.1, two structure types are supported: nanopillars and nanofins.

3.1 Common Parameters (Both Types)

  • Period
  • Height
  • Medium of the metacell

3.2 Nanopillar-Specific Parameters

  • Pillar shape: circular or quadratic
  • Diameter range and stepwidth

3.3 Nanofin-Specific Parameters

  • Fixed rectangular size
  • Rotation angle range and stepwidth
Parameters for a metasurface based on nanopillars.
Parameters for a metasurface based on nanopillars.
Parameters for a metasurface based on nanofins.
Parameters for a metasurface based on nanofins.

Important Note on Interpolation

Unlike the angle and wavelength parameters, the surrogate model does not interpolate between discrete fabrication parameter values (pillar diameters or rotation angles). Only the explicitly defined values are available during the metalens design process.

4. Train the Model

Once all parameters are configured you can start the training process. If you have access to the Distributed Computing package, training time can be significantly reduced.

  1. Start Distributed Computing as usual. (Only if Distributed Computing Package is available.)
  2. Add remote clients. (Only if Distributed Computing Package is available.)
  3. Click “Start Training”.

A progress bar and the total number of required RCWA calculations will be displayed.

Start Distributed Computing (if DC package is available).
Start Distributed Computing (if DC package is available).
Add remote clients (if DC package is available).
Add remote clients (if DC package is available).
Start training.
Start training.

Last updated: April 14, 2026 Tags: metalens surrogate model neural network optical design VirtualLab Fusion