🧭 Overview
As electrification accelerates, the need for fast, accurate, and scalable electric motor design becomes critical. Traditional modeling methods like Finite Element Analysis (FEA), while accurate, are time-consuming and computationally expensive—especially when optimizing across multiple variables like geometry, materials, and thermal conditions.
This article summarizes two AI-enhanced approaches pioneered at Altair:
- Reduced Order Modeling (ROM) using romAI
- Geometry optimization using physicsAI
Both techniques drastically reduce simulation time while maintaining high accuracy, enabling real-time applications such as Digital Twins and large-scale design space exploration.
⚙️ Part 1: AI-Driven Reduced Order Modeling (romAI)
🎯 Goal:
Accelerate simulation of induction machines while preserving high fidelity across varied operating conditions.
🧩 Background:
- Traditional approaches like the Single Phase Equivalent Circuit (SPEC) or interpolation tables are limited in scope and accuracy when voltage, frequency, and slip vary.
- romAI integrates deep learning with dynamic system modeling to create robust ROMs from a limited number of FEA simulations.
🔁 Workflow:
- Data Generation using Altair® Flux® & FluxMotor®
- 360 configurations (slip, voltage, frequency combinations)
- Outputs: Torque, losses, current, power factor, etc.
- Model Training in Altair® Twin Activate™
- High R² (>0.99) confirms strong predictive accuracy.
- ROM Application
- Deployed across >6000 new operating points with <5% error on key metrics.
- Suitable for Digital Twin integration and control-oriented simulations.
⚙️ Part 2: Geometry Optimization using physicsAI
🎯 Goal:
Optimize motor geometry (specifically for Interior Permanent Magnet Synchronous Motors) to:
- Minimize harmonic distortion of BEMF
- Minimize magnet size
- Maintain magnetic flux
🧠 AI Engine: Altair® physicsAI™
PhysicsAI is a geometric deep learning tool that learns from simulation results (mesh + physical outputs), removing the need to parameterize every geometry manually.
🔁 Workflow:
- Dataset Creation (540 DOE runs; 2 AI models)
- Model Training
- Model 1: Predict max harmonic value
- Model 2: Predict BEMF RMS
- Optimization
- Multi-objective optimization using trained AI models
- Compared to traditional meta-modeling (e.g., LSR/MLSM), physicsAI delivers faster and more accurate solutions.
📊 Results:
- Up to 15x faster iterations
- >99% accuracy for critical metrics
- Identified Pareto-optimal solutions balancing power and magnet cost
🧠 Key Takeaways
- romAI delivers fast, system-level, accurate ROMs that significantly reduce FEA costs.
- romAI can be trained with data from measurements and obtain practical system-level ROMs.
- PhysicsAI enables direct geometry-to-performance modeling, eliminating the need for manual parameterization.
- Both tools empower engineers to explore broader design spaces in less time, ideal for Digital Twin and real-time simulation scenarios.
📎 Additional Resources
- Full tutorial document is attached to this article
- For more technical details on romAI, including model architecture, training methods, and simulation parameters, Read the full IEEE publication here: AI Reduced Order Model of Induction Machine (IEEE Xplore)
- The video on Multiphysics design optimization using PhysicsAI is also attached.
