Research Bits: April 22

PIC heterogeneous integration

Researchers from Hewlett Packard Labs, Indian Institutes of Technology Madras, Microsoft Research, and University of Michigan built an AI acceleration platform based on heterogeneously integrated photonic ICs.

The PIC combines silicon photonics along with III-V compound semiconductors that functionally integrate lasers and optical amplifiers to reduce optical losses and improve scalability. The team performed lithography and dry etching on silicon-on-insulator wafers, which were then doped for metal oxide semiconductor capacitor devices and avalanche photodiodes (SPD). Silicon and germanium were selectively grown to form absorption, charge, and multiplication layers of the APD, while III-V compound semiconductors such as InP or GaAs were then integrated onto the silicon platform using die-to-wafer bonding. After adding a thin gate oxide layer, a thick dielectric layer was deposited for encapsulation and thermal stability.

According to the researchers, the platform can achieve wafer-scale integration of all the devices required to build an optical neural network on one single photonic chip, including on-chip lasers and amplifiers, high-speed photodetectors, energy-efficient modulators, and non-volatile phase shifters. [1]

Nonlinear neural network training

Researchers from University of Pennsylvania and College of Staten Island developed a programmable photonic chip that can train nonlinear neural networks using light.

The device uses a semiconductor material that responds to light. When a beam of “signal” light, which carries the input data, passes through the material, a second “pump” beam shines in from above, adjusting how the material reacts. How the signal light is absorbed, transmitted, or amplified can be controlled by changing the shape and intensity of the pump beam. This programs the chip to perform different nonlinear functions.

“Nonlinear functions are critical for training deep neural networks,” said Liang Feng, professor in Materials Science and Engineering and in Electrical and Systems Engineering at Penn, in a press release. “We’re not changing the chip’s structure. We’re using light itself to create patterns inside the material, which then reshapes how the light moves through it.”

The reconfigurable system can express a wide range of mathematical functions depending on the pump pattern, allowing the chip to adjust its behavior based on feedback from its output.

In tests, the platform achieved over 97% accuracy on a simple nonlinear decision boundary task and over 96% on the Iris flower dataset. The current work focuses on polynomials, but the team believes the approach could also enable exponential or inverse functions. [2]

Acoustically-mediated activation function

Researchers from the Max Planck Institute for the Science of Light (MPL), Leibniz University Hannover, and Massachusetts Institute of Technology (MIT) demonstrated an all-optically controlled activation function based on traveling sound waves that is suitable for a range of optical neural network approaches and allows for operation in the synthetic frequency dimension.

“The long-term prospect of creating more energy efficient optical neural networks depends on whether we are able to scale up the physical computing systems, a process potentially facilitated by a photonic activation function,” said Birgit Stiller, head of the Quantum Optoacoustics research group at MPL, in a statement.

In using sound waves as the mediator for a photonic activation function, the optical information does not have to leave the optical domain and is directly processed in optical fibers or photonic waveguides. Via the effect of stimulated Brillouin scattering, the optical input information undergoes a nonlinear change depending on the level of optical intensity.

“Our photonic activation function can be tuned in a versatile way: we show the implementation of a sigmoid, ReLU and quadratic function and the concept also allows for more exotic functions on demand, if needed for certain types of tasks,” added Grigorii Slinkov, a research associate at MPL, in a statement. [3]

References

[1] Tossoun, B., Xiao, X., Cheung, S., et al. “Large-Scale Integrated Photonic Device Platform for Energy-Efficient AI/ML Accelerators,” in IEEE Journal of Selected Topics in Quantum Electronics, vol. 31, no. 3: AI/ML Integrated Opto-electronics, pp. 1-26, May-June 2025, Art no. 8200326. https://doi.org/10.1109/JSTQE.2025.3527904

[2] Wu, T., Li, Y., Ge, L. et al. Field-programmable photonic nonlinearity. Nat. Photon. (2025). https://doi.org/10.1038/s41566-025-01660-x

[3] Slinkov, G., Becker, S., Englund, D., and Stiller, B. “All-optical nonlinear activation function based on stimulated Brillouin scattering” Nanophotonics, 2025. https://doi.org/10.1515/nanoph-2024-0513