Machine Learning, Neural Net Drive Collimated, On-Chip Nano Light-Beam Design

What you'll learn:

What prompted the MIT team to research and develop a nanoscale "flashlight"?
Details on the metasurface that was designed to create the source-beam collimator.
How a DNN, CNN, and machine learning were applied in the design process.
Results obtained through testing.

There’s significant interest in on-chip miniaturization of spectroscopic sensing arrangements used in science, biology, and medicine to implement fluorescence imaging/microscopy, infrared (IR) spectroscopy, and Raman spectroscopy (Fig. 1). All are used for quick, accurate, and non-intrusive analysis of gases and liquids.

1. Optofluidic lab on the chip sensing system. (a) Photonic waveguide technology integrated with a microfluidic chip to perform high-throughput detection and sensing. (b) Planar photonic-waveguide-based light excitation and detection for fluorescence spectroscopy. The excitation of the sample (shown in red) generates a fluorescent signal (shown in green) detected by adjoining detectors. (c) Planar waveguide-based light excitation for on-chip Raman spectroscopy. The Raman signals are represented in blue. (d) Planar waveguide-based light excitation for on-chip IR spectroscopy where excitation and detection signals are shown in red.

However, progress has been limited to some extent by lack of suitable light sources, as most of the research has centered on developing on-chip optical detectors while relatively little attention is given to the complementary excitation sources. A large part of the challenge is that the source illumination must be collimated and have a narrow wavelength spectrum, and often be tunable. Consequently, these “single-chip” analyzers still need an external LED or laser as a source, which adds to their complexity, size, and cost.

To overcome this barrier, an MIT-based a research team devised and tested a powerful nanoscale “flashlight” on a chip that produces the desired collimated light stream at a specified wavelength. While the fabrication itself used standard microelectronics processes—a big plus, of course—the design approach they used to decide what structures to build on-chip wasn’t standard.

Rather than propose design solutions and then model, simulate, and fine-tune them, they relied extensively on machine-learning techniques. The computer is taught to predict potential solutions using huge amounts of data and the physics of light propagation and manipulation.

They designed a metasurface to create a planar integrated-photonic source-beam collimator for use in on-chip optofluidic sensing applications. Further, to optimize the metasurface to better reach the various performance goals, they used the gradient-descent method with an iterative inverse-design approach.

The team started with the end result of the desired diffraction-beam profile, then applied an inverse model to evaluate the optimal design parameters of the metasurface (Fig. 2). Parameters such as the surface repetition period and height and size of scatterers were calculated using a feedforward deep neural network (DNN) and convolutional neural network (CNN) architecture.

$2. (a) Feedforward neural-network architecture to estimate the inverse model for metastructure. Diffraction profile of the metasurface is input to the network, with a four-layer architecture that has a decreasing number of units. The output of the network is eight design parameters of the metastructure. (b) Convolutional neural network (CNN) is used to report the estimation values; the approach used a single convolution layer with three fully connected layers to output eight design parameters.$

2. (a) Feedforward neural-network architecture to estimate the inverse model for metastructure. Diffraction profile of the metasurface is input to the network, with a four-layer architecture that has a decreasing number of units. The output of the network is eight design parameters of the metastructure. (b) Convolutional neural network (CNN) is used to report the estimation values; the approach used a single convolution layer with three fully connected layers to output eight design parameters.

There’s more to this approach than simply defining objectives and having the machine-learning algorithm go through options in an attempt to maximize performance along the desired directions. They note that the gradient-descent method is highly sensitive to initial conditions, so it was critical to initialize the process as accurately as possible. This was done using an effective mirror model for the meta grating structure and a cascaded mirror model with grating grooves (Fig. 3).

3. (a) Meta-grating structure with rectangular scatterers in each grating row. (b) The rectangular scatterers are dispersed within a row and their characteristics are determined by the duty cycle (C) and their period (Λx, Λy). Each row has a different characteristic of rectangular scatterers that determine how much they diffract the light off the plane to illuminate the sample. (c) Light propagation in the metasurface is easily understood through an effective mirror model where the light gets diffraction (d), absorbed (s) and transmitted (t) through the individual row.

The project wasn’t just a theoretical exercise in the use of deep machine learning to devise innovative integrated, high-performance photonic structures. Once an optimal design that met the criteria was established, they fabricated a device using single-layer patterning on a silicon-nitride substrate using e-beam lithography and reactive ion etching to define the geometry of the grating structures.

For the test and evaluation phase, the team used a beam profiler and near-infrared (NIR) camera as well as a separate alignment camera to determine the beam dispersion profile (Fig. 4). That was then compared to a conventional beam grating of comparable size (Fig. 5).

4. The experimental characterization workbench setup for the photonic device characterization. (a) Schematic of the setup. (b) Experimental setup with NIR and visible microscopic cameras to monitor the device performance.

5. Experimental comparison of metasurface and conventional binary grating-based diffractor. (a) The fabricated sample. (b,e) Microscopic top view of the light diffraction from the photonic metasurface and binary gratings. (c,d,f,g) 3D and side view of the beam profiles from the metasurface and the binary grating structure.

Obviously, performance of a device such as this can’t be summarized by a single parameter, nor is it easy to characterize it and then compare it to conventional beam-grating approaches. The researchers note that power delivered through the metasurface was approximately five times that of the conventional grating surface and the 3-dB width of the beam spot increased from 100 μm to ~300 μm with the metasurface, for footprint of 200 × 50 μm². This enables greater light illumination efficiency of the on-chip device.

The team concludes that this type of excitation source, when integrated with a planar waveguide-based photonic detector/photodetector, has the potential to miniaturize optical spectroscopy and accelerate medical sensing. Due to the dual aspects of the project—the device itself and its machine-learning design approach—the work is described in two papers published in Nature Scientific Reports, “Inverse design of photonic meta‑structure for beam collimation in on‑chip sensing” and “Design of optical meta‑structures with applications to beam engineering using deep learning.” The papers discuss all aspects of the project, from theory to conception, through execution, ML flowcharts, actual fabrication and test, along with the requisite optical physics and associated equations.