NUS Milestone: Passive All-Optical Nonlinear Neuron Activation via PPLN Waveguides Advances Photonic Computing

The NUS Breakthrough in All-Optical Nonlinear Neuron Activation

In a groundbreaking advancement for photonic computing, researchers at the National University of Singapore (NUS) have demonstrated a passive all-optical nonlinear neuron activation function using periodically poled lithium niobate (PPLN) nanophotonic waveguides. Published in the prestigious journal eLight on March 5, 2026, this work led by Professors Aaron Danner and Di Zhu addresses a critical bottleneck in optical neural networks (ONNs): the lack of efficient, passive nonlinear activations. Traditional electronic neural networks struggle with energy consumption and speed as AI models grow more complex, but photonic systems promise light-speed processing and massive parallelism. This innovation enables computations entirely in the optical domain, without electro-optic conversions that introduce latency and power overhead.

The device leverages pump-depleted second-harmonic generation (SHG) in thin-film lithium niobate (TFLN) waveguides, achieving up to 80% conversion efficiency with femtosecond response times and over 100 GHz bandwidth. This sigmoid-like nonlinearity mimics the activation functions essential for deep learning, paving the way for scalable, energy-efficient photonic AI hardware.

Why Photonic Computing is the Future of AI

Artificial intelligence demands ever-increasing computational power, with data centers projected to consume as much energy as entire countries by 2030. Photonic integrated circuits (PICs) excel in linear operations like matrix multiplications via Mach-Zehnder interferometers (MZIs) and wavelength-division multiplexing (WDM), but nonlinear activations have relied on slow, power-hungry methods. NUS's PPLN approach provides a passive solution, triggered solely by the input signal light—no external pumps, heaters, or electronics required.

Professors Danner and Zhu emphasize, "What makes this device distinctive is that the nonlinearity is triggered directly by the signal light itself." This ultrafast electronic polarization response matches the speed of linear photonics, enabling true all-optical neural networks capable of tasks like medical image classification and regression with performance rivaling digital systems.

How PPLN Waveguides Enable Nonlinear Activation: A Step-by-Step Explanation

Periodically poled lithium niobate (PPLN) waveguides are fabricated on magnesium-doped TFLN chips, where electric poling inverts crystal domains to achieve quasi-phase matching for SHG. Here's the process:

• Input Signal: Fundamental harmonic (FH) light at ~1552 nm enters the 11-mm ridge waveguide, strongly confined for high intensity.
• Pump-Depleted SHG: As power increases, χ² nonlinearity generates second harmonic (SH) at ~776 nm. Depletion of FH creates a sigmoid curve: low power yields linear response, high power saturates due to reduced pump.
• Passive Operation: No auxiliary beams; self-triggered by input, with onset at 2 mW CW or 0.02 nJ pulsed.
• Output: SH or remaining FH encodes nonlinearity, cascadable with silicon PICs for full neurons.

Experimental characterization showed 78.5% CW efficiency, 80.9% pulsed, and 1.88-2.2 dB loss—state-of-the-art for waveguide passives.

Experimental Results and Real-World Performance

The NUS team cascaded PPLN with a silicon MZI PIC for linear matrix-vector multiplications. Time-multiplexed inputs simulated 32-neuron hidden layers. On binary classification (Moon, Circle datasets), AUC reached 0.99; Iris dataset accuracy 96.7%. Multi-layer ONNs hit 82.64% on DermaMNIST-C (dermatology images) and R² 0.94 on NASA airfoil noise regression—matching PyTorch-trained digital models.

Bandwidth modeling predicted 146 GHz 3-dB, verified experimentally. Pulsed operation at 75 MHz repetition highlighted energy efficiency >1 TOPS/W potential.

Photo by Markus Winkler on Unsplash

NUS Researchers Leading the Charge

Professor Aaron Danner's Optical Device Research Group at NUS ECE focuses on chip-scale photonics with lithium niobate and barium titanate for quantum optics and neural networks. Professor Di Zhu's Integrated Quantum Photonics Lab develops single-photon tech on TFLN platforms, with expertise in nanofabrication and nonlinear optics. Co-first authors Wujie Fu and Xiaodong Shi, along with collaborators from A*STAR's IME and Q.InC, drove fabrication and testing. Their synergy exemplifies NUS's interdisciplinary strength.Rate professors like Danner and Zhu to guide peers.

Singapore's Thriving Photonics Ecosystem

Singapore invests heavily in photonics via RIE2030 ($37B), with NUS, NTU, and A*STAR leading. Events like Photonics@SG 2026 highlight integrated photonics for AI and quantum. Collaborations with GlobalFoundries advance silicon photonics foundries. This NUS work aligns with national goals for sustainable AI hardware, positioning Singapore as Asia's photonics hub.Explore Singapore higher ed opportunities.

• A*STAR's IME: Silicon nitride/LiNbO3 device integration.
• NSTIC: Semiconductor-photonics translation.
• Quantum Strategy: $37B by 2030, including photonic quantum computing.

Implications for Energy-Efficient AI and Beyond

Photonic computing could slash AI energy use—current models like GPT-4 queries consume 10x a web search; photonics offers 10-100x efficiency gains. NUS's >80% efficiency and fs speeds enable TOPS/W scales unattainable electronically. Applications span climate modeling, drug discovery, autonomous systems. Challenges like chip integration persist, but monolithic TFLN looms.NUS press release details impacts.

Comparisons with State-of-the-Art and Roadblocks Overcome

• Vs. Optoelectronic: No O/E conversions; fully passive.
• Vs. Other Nonlinearities: Faster (<100 fs) than saturable absorbers; higher bandwidth than phase-change; waveguide-integrated unlike free-space.
• Energy: Pulsed ops at fJ/neuron, vs. pJ electronic.

Overcame GVM via periodic poling; low loss via precise etching. Future: Microrings for compactness, WDM for parallelism.

Photo by Wendy Tan on Unsplash

Career Opportunities in Singapore's Photonics Sector

NUS and A*STAR offer PhD/MS programs in photonics (e.g., MSc Physics for Technology). Jobs abound: Research Engineers at IME for LiNbO3, Scientists in photonic computing. With Photonics@SG 2026 and quantum push, demand surges.Browse higher ed jobs or university positions in photonics. Craft your academic CV for success.

Future Outlook: Scalable Photonic AI at NUS

Researchers envision multi-layer cascaded chips with in-situ training. Singapore's ecosystem—NUS labs, A*STAR fabs—accelerates translation. As AI energy crises loom, this positions NUS globally. Danner and Zhu note, "This could accelerate large-scale photonic neural systems." Explore professor ratings, higher ed jobs, career advice, or post a job.