# Enhancing Model Accuracy by Replacing Lookup Maps with Machine Learning Models (Machine Learning Blog Part 1)

Written by Ryan Dudgeon

April 28, 2023## Machine Learning and Modeling Simulation

Machine learning [ML] models, such as neural networks and other types of metamodels, are fast-executing mathematical representations of data that serve a variety of modeling and simulation purposes, including:

- Replacing computationally expensive physics-based sub-systems in integrated simulation models (for instance, we are using GT-SUITE simulation models for the HVACR industry)
- Utilizing a fast surrogate model in a hardware-constrained platform such as hardware-in-the-loop (HiL)
- As an optimization approach, particularly for computationally expensive models where it might be impractical given the number of design iterations needed

## Machine Learning Models are Used in GT-SUITE

In this blog post, we’ll focus on another common situation in which ML contributes to modeling, which revolves around the use of lookup tables and maps. Depending on the application, GT-SUITE models can make extensive use of lookup tables and maps. Here they are often constructed from measurement data but can also consist of simulation data from other models. During simulations in which lookup tables and maps are used, they are usually evaluated using multivariate linear interpolation, which can be adequate when the relationship between input and output variables is linear. However, linearity is not the norm, in these situations, and linear interpolation can generate large errors when evaluating the table or map between sampled points. On the other hand, ML models are adept at capturing nonlinearities in datasets and making accurate predictions between sampled points. In GT-SUITE, neural networks and Kriging metamodels are good candidates for replacing lookup tables and maps for more accurate models.

For demonstration, consider a detailed lubrication bearing model which takes as inputs oil temperature, rotational speed, upstream oil pressure, and radial clearance. The bearing model will be run through a design of experiments (DOE) that varies these four inputs, and the predicted oil flow rate and predicted power consumption will be trained to neural networks so that the networks can be used in faster Mean Value models.

The four inputs are varied in a full-factorial DOE with 5 or 6 levels per input, yielding a total of 1080 samples. After running the 1080 simulations, we’ll have the equivalent of a lookup map for evaluating the oil flow and power consumption at any combination of the four inputs.

The 1080-sample dataset was then trained to a neural network in GT-POST consisting of two hidden layers with eight neurons each. To evaluate the predictive accuracy of the neural network in comparison to a lookup map that relies on linear interpolation, a second DOE was configured and run on the detailed bearing model, this time with 200 randomly chosen (via Latin Hypercube sampling) combinations of the four inputs.

For the 200 validation samples, the values of the four inputs were fed into both the neural network and a linear interpolating lookup map, and the predicted oil flows and power consumptions were recorded. Plots of predicted outputs vs. actual outputs are presented below.

As can be seen with the red points, linear interpolation tends to overpredict both flow rate and power. In contrast, the blue points lie on the ideal slope=1 line, showing that the neural network provides almost perfect predictions for flow rate and power.

**Learn More About Our Machine Learning Simulation Solutions**

The next time a lookup table or lookup map is needed for your GT-SUITE model, consider adding more accuracy to the model by training the data to a neural network or other ML model. For a GT-SUITE model that already utilizes one or more lookups, one might also consider upgrading it with ML.

If you are interested in learning more about how you can implement machine learning in GT-SUITE, see our productivity abilities. You can also contact us here.