What is a Reduced-Order Model?

A reduced-order model (ROM) is an accurate surrogate of a high fidelity model. Based on intelligent design space sampling and robust data fitting, the ROM acts as a statistical emulator constructed from high resolution simulation and provides accurate thermal analysis results in near real-time.

How long does it take to create a ROM?

This all depends on the number of inputs that are included in your ROM, and the amount of time that it takes to perform one simulation of your Thermal Desktop® model. A standard rule of thumb is that it typically takes on the order of 2^n number of Thermal Desktop runs that need to be performed to create an accurate ROM. E.g. You are building a ROM with 6 input factors from a TD model that takes 20 minutes to perform one run, you would want to start with 2^6 = 64 runs*20 minutes = 21 hours and 20 minutes of automated run time to create your ROM.

What can be an input or output of a ROM?

An input factor for a Veritrek ROM can be anything that you are able to control with either a symbol or a case set in your Thermal Desktop model. Outputs can be node temperatures, node group temperatures, node incident heat values, node group incident heat values, or registers.

How many inputs or outputs can you include in a ROM?

The answer to this comes down to how much time you want to spend creating your ROM. The number of training data points needed to create an accurate ROMs grows exponentially based on the number of inputs you have, but our customers typically work with up to 12 input factors. The number of output responses has no weight in terms of amount of training data needed, but the time to data-fit grows linearly as the number of output responses increases. Our customers have worked with up to 70 output responses in a reduced-order model. See some examples of our customer’s ROM metrics below.

What level of accuracy can you achieve with a ROM?

The answer to this also comes down to the amount of time you want to spend on ROM creation. Given enough time and enough training data, Veritrek can be as accurate as you need it to be. The balance comes with knowing what’s accurate enough for your specific application. We provide ROM testing within the software. Random test runs are generated, performed using Thermal Desktop and labeled as the “truth results”, and then compared against Veritrek’s predictions. Examples of test data plots that Veritrek creates for you are shown below with some customer examples. We want the test data to fall along the ideal, diagonal dashed line that represents a perfect ROM prediction. However, we can see some cases where the ROM prediction is not exactly equal to the truth result. This is where the user must make the judgement call on if the ROM is accurate enough, or they can add more training data on top of your existing ROM to improve its accuracy.

Does Veritrek work with transient simulations?

Veritrek works with both steady-state and transient simulations. For transient simulations, Veritrek can pull either the minimum, mean, and/or maximum output value during the transient analysis.