The panellists were divided in 2 teams:
AI Team
Alfonso Martínez – Frenetic
Minjie Chen – Princeton University
Joao Pinto – Oakridge National Labs
Conventional Team
Alex Huang – University of Texas at Austin
Rolando Burgos – Virginia Tech
Dragan Maksimovic – University of Colorado Boulder
As you can see, too many Academics, but…such a great academics! 🎓 Frenetic was there, because probably is the one the closest companies to the state of the art of the industry.
The first debate represented very well the baby status:
1. It would be an AI able to design a Magnetic better than the conventional method?
The question itself shows how far we are from understanding the mission of the AI.
The mission of the AI (let’s call it, Machine Learning instead) is not to replace a human designing a complete system, but solving specific problems, which our understanding is very poor or the existing models are very bad. We are very far of asking a ML system to design a complete magnetic. However, we are breaking the magnetic in different calculations and problems and analyzing which of them could be solved with ML, because the current model doesn´t have sense.
For example, a ML wouldn´t be useful for calculating the DC resistance, because the current model we have is very good. We understand very well that model and the equation is very simple, therefore, wouldn’t have sense to create a ML model.
However, Leakage inductance or High frequency resistance models are very poor. The current method of learning is trial and error iterations. Therefore, these are the parameters where we can measure parameters and create ML models to give us very fast responses.
For example, at Frenetic, we have a very accurate Machine Learning model of Leakage inductance, who has been trained combining measurements from the lab and simulations.
2. Is there any distinction between good data and bad data?
The answer to this question was very fun, because first, we would need to define what means good and bad data. The quality of the data is very important, because the model learn from them, Alfonso quotes during the RAP session:
Having wrong data is like having a bad teacher
All the panellists agreed the importance of having great data for building AI models.
At the end of the Rap session, Dragan Maksimovik said:
I want to change to ML team (😊).
He also said, there are complex problems, where using ML is our best option nowadays as there are other problems, we understand very well and a conventional method is good enough.
After the session, I thought, we need to explain what the most relevant people are doing about ML in Power Electronics and especially in Magnetics. Therefore, we are going to include a module of AI in our Training Program.
Would you like to know more about how we build models of AI ? Join the Training,
You prefer to attend next APEC with zero idea about AI ? Not need the training 😉.
The Frenetic HF Magnetic Training Program sessions are:
- Module 1: Full Bridge Phase Shift Transformer
- Module 2: PFC Inductor
- Module 3: Magnetics Industrialization
- Module 4: Lab Measurements with the Climatic Chamber
- Module 5: Machine Learning in Magnetics: Why, Where and How
- Project Assignment: You will design a magnetic during the training program
