For many years, physicists used their minds to make major breakthroughs and advances in their field. What if artificial Intelligence could aid in these discoveries?
According to a press release from the university, Duke University researchers demonstrated last month that machine learning algorithms can incorporate known physics. The team created a machine-learning algorithm that could deduce the properties and interact with electromagnetic fields. This was a first-of its kind project.
Predicting metamaterial properties
These results were extraordinary. The algorithm was able to predict the metamaterial’s properties better than any previous method and also provided new insights.
Professor of electrical and computer engineering at Duke, Willie Padilla said, “By incorporating existing physics directly into machine learning, the algorithm is able to find solutions with fewer training data and in a shorter time.” While this study was primarily a demonstration that the approach can recreate known solutions it also provided some insight into the inner workings non-metallic metamaterials.
The researchers were focused on finding new insights that were both accurate and logical in their research.
Jordan Malof, an assistant professor of electrical and computer engineering at Duke, stated that neural networks attempt to find patterns in data but sometimes they don’t follow the laws of Physics, making the model it creates unpredictable. We made it impossible for the neural network not to find relationships that might fit the data, but aren’t true by forcing it to follow the laws of Physics.”
This was done by imposing a Lorentz model onto the neural network. This equation describes how an electromagnetic field interacts with the intrinsic properties of a material. However, this was not an easy feat.
Omar Khatib is a postdoctoral researcher in Padilla’s lab. He said that “when you make a neural net more interpretable, which was in some sense what they’ve done here,” it can be more difficult to fine tune. “We had difficulty optimizing the training to learn these patterns.
This model is significantly more efficient
Researchers were surprised to discover that the model performed better than the previous neural networks they had developed for similar tasks. The model required significantly fewer parameters to determine metamaterial properties. The model can even discover new properties on its own.
The researchers are now ready to apply their approach to uncharted territory.
Padilla stated, “Now that we have shown that this is possible, we want this approach to systems where physics remains unknown.”
Malof said that while many people use neural networks to predict the material properties of materials, it is difficult to get enough training data for simulations. This work shows that models don’t require as much data and is therefore useful for everyone.