There are several applications for composite materials with two or more different components in contemporary science and technology. The shrinking of lenses, lasers, and detectors – the fundamental technologies that underpin telecommunications, imaging, and sensing — is one of their most significant applications in optics, where these materials enable future miniaturization.
To identify an optical composite structure suitable for a specific situation, scientists must understand how light travels through the material. However, solving the equations of optics in a multilayer media is both analytically and numerically challenging. In a recent paper published in Advanced Photonics Research, researchers devised a system based on machine learning that can deal with challenges of this nature extremely well by calculating the light ray paths in a specific composite material.
How to teach machines
Machine learning is a type of data analysis that enables software to find patterns by learning from input data. In recent years, machine learning has become an integral aspect of our lives, powering picture identification, automated translation, self-driving vehicles, and even forecasting battery performance and medication development, etc. In the majority of these applications, an algorithm is trained on vast quantities of “labeled data” – millions of photos including tagged people, automobiles, bicycles, or whole books annotated with their translations.
“Machine learning may also be utilized in the scientific realm,” said Viktor Podolskiy of the University of Massachusetts Lowell, one of the study’s authors. “In this situation, it is customary to forecast the answer to an equation or the outcome of an experiment. Typically, machine learning must be taught using a library of known answers or the outcomes of previous experiments in order to be utilized effectively. Given sufficient past data, machine learning appears to perform well.
However, in this usual situation, known as “black box” machine learning, researchers ignore the scientific knowledge that mankind has gathered over the millennia and train the computer using just input/result pairings. In a sense, the physicists’ grasp of how to derive and solve the governing equations goes to waste. In cases where brute force solutions are challenging, there are therefore insufficient data to train machine learning.
Bringing composite materials to light
The challenge of light propagation in multilayer materials with varying optical characteristics necessitates precisely this type of complex computation. To address this issue, the authors of the paper devised physics-informed machine learning, which augments conventional “black box” algorithms with the known electromagnetic field dynamics equations that control light propagation.
Podolskiy noted, “Our objective was to leverage some of the ‘extra’ scientific knowledge in the training process, combining the benefits of science with machine learning.”
Light transmission in a ten-layer optical composite was the model problem being solved by the scientists. Using a dataset consisting of light paths in hundreds of composites with varying optical characteristics, they trained both the “black box” and physics-informed machine learning methods.
It found out that the approach they devised was far more efficient and used a dataset 20 times less than standard machine learning to get the same prediction accuracy. An even more striking benefit over a numerical solution was that trained machine learning algorithms discovered the combinations of light beams hundreds of times quicker.
Podolskiy continued, “Our results imply that the introduction of scientific information into the training process helps us to train models more quickly, with a significantly lower quantity of training data, and to build models that can function correctly with a much larger variety of input parameters.”
A promising future
The researchers anticipate that their shown method can be extended to additional optical difficulties. They also anticipate that their approach might be enhanced and made much more potent.
Podolskiy said, “We are presently attempting to expand our methodologies to a broader class of situations in order to establish a ‘hybrid’ framework that would accelerate the study of light interaction with composites.” “The strategy would initially apply time-consuming science-based solvers for a few of data points, and then use this early data as a training set for considerably quicker physics-informed machine learning models. We anticipate that the new framework will let us to do computations that previously required weeks in a matter of days or even hours.”