COMSOL Multiphysics^® 6.4 Release Highlights

COMSOL Multiphysics^® version 6.4 introduces enhanced capabilities for surrogate models, including the ability to export trained deep neural network (DNN) functions as well as new support for batch and cluster computing. Learn more about these updates below.

DNN Export

In this release, COMSOL Multiphysics^® introduces support for exporting trained DNNs to the open ONNX format. This update enables the use of DNNs in MATLAB^® and Simulink^® as well as other external tools, allowing seamless integration of COMSOL Multiphysics-trained DNNs into external workflows and a range of other machine learning environments.

The Export feature used to export the DNN function in the ONNX format.

Batch and Cluster Support for Surrogate Model Training Data Generation

Batch and cluster computing are now supported with the Surrogate Model Training study, enabling efficient parallel computation for large training datasets. Using the Batch study step with surrogate model training, it is possible to automatically distribute multiple simulations across available cores. By using the Cluster Computing study step with a Surrogate Model Training study, these tasks can be executed simultaneously on a cluster, significantly reducing total computation time for data generation. This enhancement provides the ability to run multiple simulations in parallel while maintaining complete control over results storage and data synchronization.

A Surrogate Model Training study is combined with the Cluster Computing study step to distribute simulation runs across a cluster and automatically synchronize the generated training data.

New Tutorial Models

COMSOL Multiphysics^® version 6.4 offers two new surrogate modeling tutorial models.

Microstrip Patch Antenna Surrogate

This example demonstrates training and using a DNN surrogate model to rapidly estimate the performance of a microstrip patch antenna based on four design parameters: patch length, tuning stub length, substrate dielectric constant, and operating frequency.

Application Library Title:
microstrip_patch_antenna_surrogate
Download from the Application Gallery

*Requires the RF Module

Modeling Space-Dependent Plasmas with Deep Neural Network Surrogate Models

Electron transport in low-temperature plasmas depends strongly on the electron energy distribution function, often non-Maxwellian in reality. This model uses a DNN trained on Boltzmann equation solutions to include kinetic effects in fluid simulations with much lower computational cost.

Download from the Application Gallery

*Requires the Plasma Module and the AC/DC Module