&Bullet; physics 14, 90

A machine learning approach quickly characterizes an optical fiber and identifies transmission channels that are not affected by deformation.

Fiber marker. Rays of spatially modulated light are sent through a multimode fiber. The recorded output light is fed into a machine learning algorithm that determines how each mode or data channel in the fiber is scrambled by bending and other perturbations.Fiber marker. Rays of spatially modulated light are sent through a multimode fiber. The recorded output light is fed into a machine learning algorithm that determines how each mode or data channel in the fiber is processed by bending and other … show more

Optical multimode fibers, which can transmit many simultaneous signals, offer a possibility to increase the data rates beyond those of commonly used single-mode fibers. However, they are extremely sensitive to environmental disturbances that scramble the signals. Researchers have now developed a numerical approach to quickly assess this interference and identify a number of signal channels in the fiber that are resistant to interference [1] . Using such robust channels would improve the performance of multimode fibers, which could increase transmission rates in communication networks and lead to less invasive endoscopic probes for imaging body organs.

Light in a multimode fiber does not propagate in the middle but in a zigzag path with reflections from the outer cylindrical shell of the fiber. Several zigzag paths or modes are available in the fiber, each of which serves as a separate data channel. The modes can be distinguished by the spatial pattern of bright spots that each creates when projected directly onto a screen.

In a perfect fiber with a fixed temperature, a light signal fed in in one mode emerges unchanged in the same mode at the end of the fiber. But manufacturing defects, temperature changes, and fiber deformations can cause coupling between modes, allowing light to “leak” randomly from one mode to another. Researchers characterize this mode-to-mode coupling with a mathematical unit, the so-called transmission matrix. Obtaining a reliable transmission matrix requires very accurate optical measurements which must be repeated every time the fiber is bent or otherwise disturbed.

Sébastien Popoff from ESPCI Paris and his colleagues have now used tools from machine learning frameworks to develop an approach to quickly measure the transmission matrix and then identify robust channels that are not very sensitive to changes in the fiber. The researchers first generated light beams with brightness variations across their profiles. They then injected these rays into a 30 cm long fiber and imaged the light emerging at the other end with a 1 kHz frame rate camera. After about 10 seconds, they received a transmission matrix that provides the relationship between the input light and the output light at each spatial location or “pixel” within the beam.

S. Popoff / ESPCI Paris

Data squeezing. Some light modes that are fed into an optical fiber remain robust against deformations that are introduced by a pin pressing on the fiber cladding.

However, the pixel-to-pixel transfer matrix must be converted to the mode-to-mode transfer matrix in order to examine signal interruption. Such a conversion is made difficult by slight misalignments within the fiber and camera system. Instead of trying to measure the misalignment, the researchers developed a numerical model that mimicked the effects of the misalignment. Using an optimization process, the model corrects the misalignments and returns the desired mode-to-mode transmission matrix for a total of 110 modes within a few seconds. “We’re effectively shifting the complexity from experimental setup to numerical optimization,” says Popoff.

Using a method that can calculate the transmission matrix in around 10 s, the researchers were able to study the effects of the deformation. In a series of experiments, they used a pen to press differently on the fiber, repeating the transmission matrix measurement each time. They found that the deformation drastically affected the propagation of light in the fiber, but the impact was “localized” in the sense that light emitted into one mode was only coupled to certain other modes, even with severe deformations.

The researchers then used their knowledge of the transmission matrix and its sensitivity to certain fiber deformations to select a series of channels that would remain insensitive to deformations. Choosing the most robust channels for use in multimode fiber could ensure that independent signals sent on each of those channels do not mix with others, even if the configuration of the disturbance changes. “If I know that my fiber is always bent and a channel is insensitive to this type of deformation, then that channel is fine, even if the fiber bending changes over time,” says Popoff.

Tomáš Čižmár, fiber expert at the Leibniz Institute for Photonic Technologies, says the new method shortens the time to determine the mode-to-mode transmission matrix. If the technique can be validated for a wider range of deformation types, “it could stimulate numerous applications, including fiber-based endomicroscopy,” he says.

–Rachel Berkowitz

Rachel Berkowitz is the corresponding editor for physics based in Vancouver, Canada.


  1. M. Matthes et al., “Learning and avoiding interference in multimode fibers”, Phys. Rev. X11, 021060 (2021).

areas of expertise

On the subject of matching items

Pulses of light change the speed in a plasma
Bring high resolution x-ray imaging to the laboratory
Make heart cells glow with tiny lasers

More articles


Please enter your comment!
Please enter your name here