Neural Networks and Precision Tapering Bring Photonic Lanterns Into Focus

A photonic lantern sounds like something from science fiction, but it is one of the most elegant devices in modern fiber optics: a bundle of ordinary single-mode fibers that gradually fuse and taper into a single multimode fiber, allowing light to morph smoothly between discrete spatial modes without scattering away. First proposed in 2005, the device has become a quiet workhorse behind high-power fiber lasers, next-generation telecommunications, and even astronomical telescopes that hunt for Earth-like planets. Yet engineers have long struggled with two stubborn problems. Simulating how light actually propagates through a lantern can take hours per run, and fabricating one with low loss demands near-impossible control over heat, cleanliness, and alignment. A new systematic study published in the journal iScience by Yao Lu, Zilun Chen, Tong Liu, Zhuruixiang Sun, Xiang Li, and Zongfu Jiang tackles both challenges head-on, describing a complete chain from theoretical modeling through fabrication to real-world performance testing of working devices.

The heart of the difficulty lies in the physics of the taper. A photonic lantern is a non-uniform waveguide, meaning its refractive index distribution changes along its length, and there is no clean analytical solution for how the light field evolves inside it. Physicists therefore rely on numerical methods. The team first turned to the beam propagation method, or BPM, a classic technique dating back to 1978 that steps the optical field forward through the device slice by slice. Using a five-by-one lantern built from single-mode fibers with 10-micrometer cores feeding a 30-micrometer-core multimode output, the simulations revealed the full story of mode evolution: five separate fundamental modes merge into collective supermodes inside the taper, then emerge as a broad guided mode in the multimode fiber. But the method comes at a steep price. Each simulation took one and a half to two hours, errors accumulated over distance, and every new input field required a fresh calculation, making batch design optimization or real-time adaptive control effectively impossible.

To break that bottleneck, the researchers adopted the mode expansion propagation method, which divides the waveguide into many short sections, solves for the eigenmodes of each section using Maxwell’s equations, and stitches the sections together through overlap integrals. This approach yields both forward and backward transmission matrices at once, keeps errors from compounding, and crucially allows any new input field to be computed by simple matrix multiplication once the transmission matrix is known. For a three-by-one mode-selective lantern, the method produced detailed field distributions, mode-dependent loss curves, and a concrete design rule: once the taper length exceeds roughly 17 millimeters, losses drop to negligible levels. A full lantern simulation took only about ten minutes instead of hours, a dramatic improvement that still fell short for the kind of rapid, iterative modeling that adaptive optical systems ultimately require.

The decisive leap came from machine learning. Because a photonic lantern’s behavior is a deterministic mapping between input and output light fields governed by Maxwell’s equations, it is ideally suited to supervised learning. The team built a dropout neural network whose inputs are the amplitudes and phases of each input beam and whose outputs are the complex coefficients of each output eigenmode. Each hidden layer contained 2,000 neurons with LeakyReLU activations, and dropout regularization randomly silenced 20 percent of neurons during training to prevent overfitting. Training used 1,500 experimental field pairs from a six-by-one lantern, expanded to 6,000 examples through data augmentation exploiting the symmetry of the six-fiber structure, and ran for 25,000 epochs with an adaptive learning rate schedule.

The results were striking. The best network, with three hidden layers, predicted light intensity with a mean squared error of just 9.793 times ten to the minus five and a coefficient of determination of 0.9973, while phase prediction reached an error of 3.772 times ten to the minus three. Once trained, the model produced a prediction in roughly 60 milliseconds on a standard laptop processor, an acceleration of more than five orders of magnitude compared with the hour-and-a-half beam propagation simulations. Perhaps most importantly, the approach can be trained directly on experimentally measured input-output data from a finished device, offering a way to calibrate real lanterns whose as-fabricated geometry inevitably deviates from design blueprints. The authors note that generalizing across different lantern geometries will require larger, more diverse datasets, an explicit target for future work.

Modeling was only half the battle. The team also systematically developed the fabrication of lanterns made by inserting a bundle of single-mode fibers into a low-refractive-index glass capillary and fusing the assembly while stretching it. They compared three device families: three-dimensional integrated lanterns written with ultrafast lasers, which suffer from scattering-related losses; tapered multicore fibers, which are efficient but hard to splice and difficult to address individually; and the capillary-sleeve fiber bundle approach they favored, which keeps the input fibers separated and accessible while using cheap, standard step-index fibers as raw material. Using a Vytran GPX3400 glass processing workstation and a Fujikura FSM-100P+ splicer, the researchers walked through material pretreatment, acid etching of fiber claddings, capillary pre-tapering, bundling, fusion tapering, precision cleaving, and dissimilar-fiber splicing.

The process details turned out to make or break the device. In 60 controlled tapering attempts, fabrication yield fell from about 85 percent under cleanroom conditions to roughly 35 percent with minor contamination and below 10 percent with poor drying, because residual particles ignite during fusion and become scattering centers; particles larger than 5 micrometers caused thermal runaway at just 2 watts. Bundling symmetry proved equally critical: fibers arranged in regular polygons preserve the lantern’s mode evolution capability, and unevenly distributed residual air holes introduced more than 1 decibel of loss and over 3 decibels of mode purity degradation, while symmetric holes of varying size mattered far less. The team defined an optimal collapse regime in which the capillary diameter remains about twice the circumscribed circle of the fiber bundle, achieving a splicing success rate near 90 percent. Alignment thresholds were equally strict: lateral misalignment beyond half a micrometer or tilt beyond 0.3 degrees triggered excess losses above 0.65 decibels.

Performance testing of the finished devices validated both the simulations and the process. At 1,064 nanometers with water cooling, a fabricated three-by-one non-mode-selective lantern transmitted more than 98 percent of the light when input power stayed below 1.5 watts per arm. At higher powers, efficiency dropped, falling to about 62 percent at 3 watts, a reversible degradation the team attributes to thermal effects in the taper waist rather than permanent damage; devices survived 5 watts per arm without harm. Mode field imaging showed output spot patterns matching simulations with overlap integrals exceeding 0.965, and multi-arm input tests revealed continuously shifting output spots driven by environmental phase drift, confirming that practical mode control demands an adaptive feedback loop with phase locking.

Mode-selective lanterns delivered on their promise with quantified crosstalk. In a three-by-one device with input fibers of different core diameters, injecting light into the 15-micrometer-core arm yielded a fundamental mode with 94.6 percent purity, while the two 10-micrometer arms produced LP11-like modes with purities of 78.6 and 84.3 percent. A larger six-by-one lantern selectively excited each of the first six LP mode groups, though with noticeably higher crosstalk, illustrating the real-world gap between ideal designs and finished hardware. Looking ahead, the authors propose double-clad input fibers to shorten the required adiabatic taper and photonic crystal structures arranged around the fiber bundle to suppress light leakage. Together, the modeling framework and the fabrication playbook bring photonic lanterns closer to routine deployment in kilowatt-class fiber lasers, capacity-crunching mode division multiplexing, and wavefront sensors that could sharpen humanity’s view of distant worlds.

Subject of Research: Transmission characteristics and fabrication of photonic lanterns for mode control in fiber optics

Article Title: Research on the transmission characteristics and fabrication process of photonic lanterns

Article References: Research on the transmission characteristics and fabrication process of photonic lanterns. (n.d.). https://doi.org/10.1016/j.isci.2026.117615

Image Credits: AI Generated

DOI: 10.1016/j.isci.2026.117615

Keywords: photonic lantern, fiber optics, mode division multiplexing, neural network, beam propagation method, mode expansion propagation, high-power fiber lasers, fused tapering, spatial mode control, wavefront sensing, optical waveguides, mode purity