Two materials with different physical and chemical properties combine to form a composite material. These materials find their use in various areas of modern science and technology. The significant application of composite materials in the field of optics, where they help design lenses, lasers, and detectors in small sizes. These are the fundamental technologies that support telecommunications, imaging, and sensing.
Scientists must be aware of the way light propagates in a material, in order to find an optical composite structure suitable for a given problem. However, solving the optics equations in a multilayer medium is a challenge both analytically and numerically.
This challenge can be overcome by a machine learning-based technique developed by researchers which computes the trajectories of the light rays in a specified composite material, according to the latest study published in Advanced Photonics Research.
Machine learning (ML) is a data analysis technique that enables software to learn from input data and recognize patterns. Machine learning has evolved so much that it has become a part of our everyday lives and is involved in activities such as powering image recognition, automatic translation, anticipating the performance of a battery, autonomous driving, discovery of drugs, etc.
The majority of these applications involve the algorithm being trained on massive amounts of “labeled data” which includes millions of images of people, cars, bicycles, or entire books annotated with translations.
Machine learning can also be utilized in the scientific domain, according to Viktor Podolskiy – one of the study’s authors belonging to the University of Massachusetts Lowell. In the scientific domain, it is generally applied to predict the solution to an equation or the outcome of an experiment.
Machine learning can be efficient when it is trained on the library of familiar solutions or based on the results of previous experiments. With sufficient previous information, machine learning can perform effectively.
Nevertheless, in this typical scenario known as black box” machine learning, researchers overlook the scientific knowledge gathered by humanity for centuries and train the computer using only input/result pairs. Hypothetically, the physicists’ understanding of how to derive and solve the governing equations is rendered useless.
As a result, in situations where brute force solutions are tough, there is insufficient data to train machine learning.
Physics-informed machine learning
An example of the difficult calculation is the multilayer materials where the layers contain different optical properties and there is a problem with light propagation. To solve it, the study’s authors proposed physics-informed machine learning, which augments traditional “black box” algorithms with known equations of electromagnetic field dynamics that govern light propagation.
Podolskiy explained that their motive was to use a few of the ‘extra-scientific knowledge in the training procedure, apparently combining the advantages of science and machine learning.
The physicists were working on a model of light propagation in a ten-layer optical composite. They used a dataset of light trajectories in hundreds of composites with varying optical properties to train both the “black box” and the physics-informed machine learning algorithms.
The technique developed turned out to be very much effective, requiring a dataset roughly 20 times smaller to achieve the same prediction accuracy as conventional machine learning. A significant advantage over a numerical solver was that trained machine learning algorithms found the light ray configurations hundreds of times faster.
Podolskiy further added that their findings indicate that incorporating scientific knowledge into the training process allows them in training the models quicker, on a very limited amount of training data, and develop models that work accurately on a much broader set of input parameters.
A promising future
The researchers believe that the approach they demonstrated in their work can be applied to other optics problems. They also believe that their method can be improved and made even more powerful.
Podolskiy concluded that they were currently working on expanding their approaches to a broader class of problems, intending to develop a ‘hybrid’ framework that would speed up the analysis of composite light interaction.
The approach would first use tedious science-based solvers for some data points, then use this initial data as a training set for much quicker physics-informed machine learning models. They count on the resulting framework to enable them to run calculations that previously took weeks in a matter of days or even hours.