Machine Learning-Based E-Motor Thermal Metamodeling: Replacing FE Models for Fast System-Level EV Simulation

The Challenge with E-Motor Thermal Simulation

Thermal simulation of electric motors sits at the intersection of two competing demands: the need for spatial accuracy at the component level, and the speed required for system-level scenario exploration.

High‑fidelity 3D Finite Element (FE) thermal models of electric machines are essential for predicting hotspot temperatures, validating cooling strategies, and ensuring durability under demanding real‑world conditions. However, these models are also computationally expensive, slowing down system‑level simulations where hundreds of scenarios must be evaluated quickly.

While lumped thermal mass models run much faster, they lack the spatial resolution needed to accurately capture local peak temperatures, and model calibration is typically required. To bridge this gap, Machine Learning (ML) metamodels offer a new approach based on near-instant execution and ability to retain the hotspot prediction capability.

This blog introduces a data driven method to convert a fully physical FE e-motor thermal model into a dynamic regression metamodel, combining speed, robustness, and hotspot level fidelity.

High-Fidelity FE Thermal Model of a GT-SUITE-Integrated E-Motor

The starting point is a detailed, physics‑based FE thermal model of an electric drive motor. The model captures structural heat conduction, losses from various components, coolant flow, oil spray effects, and boundary conditions driven by a vehicle‑level GT‑SUITE model.

This setup provides accurate 3D temperature distribution, enabling prediction of stator, winding, rotor, and magnet hotspots. While the FE model delivers high‑fidelity thermal insights, its 3.5x slower‑than‑real‑time performance limits the speed of repeated scenario sweeps across long drive cycles.

Thus, replacing the FE domain with an ML‑based metamodel provides a way to drastically accelerate system simulations while preserving thermal accuracy.

DOE-Based Training Data Generation from the FE Thermal Model

Robust ML performance depends heavily on the quality and diversity of training data. Here, the FE thermal model itself served as a virtual data generator.

A custom DOE was constructed to span a wide operating envelope:

• 60 transient simulation samples
• Variations across:
  • 4× driving cycles
  • 5× coolant inlet temperature levels
  • 3× ambient conditions
  • Variable coolant flow rate
  • Variable oil spray cooling

These inputs drive the nonlinear thermal behavior of the motor. The resulting FE simulations yield time‑series outputs for all relevant hotspot temperatures and coolant outlet temperature.

Selecting Thermal Inputs and Hotspot Output Signals for the Metamodel

From the DOE dataset, the following 9 input signals (factors) were selected as drivers of thermal behavior:

• Stator losses
• Winding losses
• Rotor losses
• Magnet losses
• Windage losses
• Ambient temperature
• Coolant flow rate
• Coolant inlet temperature
• Oil spray cooling (on/off)

The metamodels predict 5 key outputs (responses):

• Stator max temperature
• Winding max temperature
• Rotor max temperature
• Magnet max temperature
• Coolant outlet temperature

These signals represent the most critical thermal performance indicators for system‑level studies.

Training a NeuralODE Dynamic Regression Metamodel in GT-SUITE

GT‑SUITE’s Machine Learning Assistant enables guided metamodel development, offering dynamic regression metamodels to represent continuous‑time dynamics and capture non‑linear transient behavior. For this application, the NeuralODE metamodel type demonstrated a slightly higher accuracy.  8 full-length driving cycles were held out from the training process and reserved for testing.

Regression and transient response plots for training, validation, and test datasets showed strong agreement between neural outputs and GT‑SUITE reference data:

• Smooth, stable behavior
• Accurate prediction of hotspot temperature peaks
• Low absolute error, even during high thermal transients

 

After training, metamodels are exported to .mmp files for native GT‑SUITE integration. FMU and C code export are also available for third‑party integration or ECU deployment.

System-Level Integration and Validation Using the MetamodelHarness in GT-SUITE

Validation in the isolated Machine Learning Assistant environment isn’t enough, integration into the existing plant model is essential for a robust assessment of the metamodels behavior.

Within the system‑level model:

• The FE thermal domain is fully replaced using the MetamodelHarness template
• The metamodels predict hotspot temperatures and coolant outlet temperature
• A dedicated 1‑second timestep control circuit is used to match training conditions
• Vehicle‑level losses dynamically feed the metamodel

Different maneuvers including unseen drive cycles and altered boundary conditions were tested across:

• Physical baseline 3D FE thermal model
• Neural network metamodel

Results showed that the surrogate model accurately tracked reference behavior across all scenarios. Replacing the FE thermal model with neural networks reduced CPU load by over 99%.

Conclusion: Practical ML-Accelerated Thermal Simulation for EV Motor Development

Fully‑physical FE thermal models remain the gold standard for predicting hotspot temperatures in electric machines, yet their computational cost makes them impractical for large sets of system‑level simulations. By contrast, the data‑driven metamodel developed in this study demonstrates that it is possible to retain the fidelity of FE‑based hotspot prediction while achieving near‑instant execution.

By training dynamic regression metamodels on a carefully designed DOE generated from the FE model, the resulting metamodel maintains strong accuracy even on unseen drive cycles.

Once integrated into the GT‑SUITE system model, it delivers thermal predictions at a fraction of the computational cost and enables fast, scalable scenario exploration. This makes high‑fidelity hotspot prediction feasible for everyday system‑level studies where speed and robustness are equally essential. This approach has been validated in production simulation workflows by Gamma Technologies engineers and reflects the same methodology available to GT-SUITE users through the Machine Learning Assistant module.

Discover how integrating physics-based models with ML metamodels can significantly reduce simulation time while preserving fidelity. Read our blogs and visit our Machine Learning and Optimization Solutions page for more related content. Join our LinkedIn community to stay updated on simulation-driven workflows, battery modeling insights, and upcoming technical content. You can also contact us to learn how Gamma Technologies can support your machine learning goals.

 

 

Frequently Asked Questions

What is an e-motor thermal metamodel and how does it differ from a lumped thermal model?

An e-motor thermal metamodel is a machine learning surrogate trained on simulation data from a high-fidelity FE (Finite Element) model. Unlike a lumped thermal model which requires manual calibration, an ML metamodel learns the nonlinear transient relationships between component losses, cooling conditions, and hotspot temperatures directly from FE data. The result is a model that retains the spatial fidelity of FE hotspot prediction (stator, winding, rotor, magnet) while running at near-real-time speed, without the calibration overhead of lumped models.

 

How much does replacing the FE thermal model with an ML metamodel reduce simulation time in GT-SUITE?

In this case study, replacing the FE thermal domain with the NeuralODE metamodel reduced CPU load by over 99%. The original FE model ran at 3.5× slower than real time, making it impractical for repeated scenario sweeps. The integrated metamodel runs faster than real-time within the GT-SUITE system model, enabling fast exploration of hundreds of drive cycle and boundary condition combinations without sacrificing hotspot prediction accuracy.

 

Can the trained ML metamodel be exported and used outside of GT-SUITE, for example in an ECU or third-party tool?

Yes. Once trained and validated, GT-SUITE metamodels can be exported in multiple formats beyond the native .mmp file. FMU (Functional Mock-up Unit) export enables integration into other simulation environments. C code export supports deployment on embedded hardware, including ECUs, making these ML-based thermal predictors applicable to real-time thermal protection strategies in production vehicle control systems.