Passive All-Optical Nonlinear Neuron Activation via PPLN Nanophotonic Waveguides

Abstract

Artificial intelligence (AI) is transforming modern life, yet the growing scale of AI applications places mounting demands on computational resources, raising sustainability concerns. Photonic integrated circuits (PICs) offer a promising alternative, enabling massive parallelism, low latency, and reduced electrical overhead, particularly excelling in high-throughput linear operations. However, passive and fully optical nonlinear activation functions with equally superb performance remain rare, posing a critical bottleneck in realizing all-optical neural networks in PICs. Here, we demonstrate a compact and integrated all-optical nonlinear activation method, experimentally realized through strong second-order optical nonlinearities in periodically poled lithium niobate (PPLN) nanophotonic waveguides, achieving 80% absolute conversion efficiency. This activation exhibits a sigmoid-like, wavelength-selective response with femtosecond-scale dynamics and light-speed processing, operating passively without external control and auxiliary signals. We validate its feasibility for neural inference by cascading the PPLN-driven activations with a linear silicon PIC, demonstrating all-optical nonlinear neuron expressivity. Moreover, combining the measured nonlinearity with linear operations calculated by the PIC, we show that PPLN-activated multi-layer optical neural networks can achieve performance on par with digital implementations in real-world tasks, including airfoil regression and medical image classification. These results pave the way toward scalable, high-speed, and fully integrated all-optical neural networks for next-generation photonic AI hardware.

Figures (5)

• Figure 1: All-optical neuron architecture employing pump-depleted SHG in PPLN nanophotonic waveguides to provide on-chip nonlinear activation with MZI-based linear processing.a Microscope images of the MZI and PPLN chips. Input features and network parameters are programmed into the phase shifters through electrical pads, and intermediate predictions are extracted from the on-chip optical output ports. b Neuron inputs are encoded in the amplitudes and phases of light at the fundamental harmonic frequency $\omega$. These coherent optical signals propagate iteratively through the MZI network and the PPLN array, where dot products are executed via optical interference, and nonlinear activation is realized through efficient, passive pump-depleted SHG enabled by quasi-phase matching and strong $\chi^{(2)}$ parametric interaction from tight optical confinement in the TFLN waveguide. The neuron's output is extracted via homodyne detection at the same frequency.
• Figure 2: Pump-depleted SHG process in a PPLN nanophotonic waveguide for all-optical nonlinear neuron activation.a PPLN device fabrication and characterization. (i) Cross-sectional SEM image. (ii) Top-view laser-scanning SHG microscope imaging. (iii) Measurement setup: output light is wavelength demultiplexed by a WDM into FH and SH channels and detected by photodiodes (PD). b SHG spectrum of the fabricated PPLN waveguide, showing a phase-matching wavelength at 1552 nm. c Measured CW-pumped nonlinear activation (dots) of the FH and SH light. Solid curves represent scaled Runge–Kutta simulation, revealing a saturable, sigmoid-like nonlinearity for the FH light with a nonlinear onset threshold of approximately 2 mW and a saturation window emerging below 10 mW (inset). d Calculated absolute conversion efficiency from the FH to SH as a function of CW pump power. e Numerical SH power transfer function versus pump modulation frequency, with the 3-dB activation bandwidth set by group-velocity mismatch. f Measured pulsed-pumped nonlinear activation of the FH light and corresponding conversion efficiency, showing a nonlinear onset threshold near 0.02 nJ and a sigmoidal saturation window below 0.15 nJ.
• Figure 3: Design and characterization of the MZI system on a PIC for the linear operation required by the proposed optical neuron.a Schematic and microscopic image of the integrated silicon photonics-based MZI system, comprising on-chip multimode interferometers (MMIs), thermo-optic phase shifters, PDs, and wire bonding for electrical control. The system supports arbitrary $2\times2$ matrix–vector multiplication. b Experimental characterization of the PIC-based optical matrix multiplication using randomly generated input matrices to evaluate linear computation fidelity. The measured outputs yield a standard deviation of 0.005 and an $R^2$ of 97.7% under CW operation, and 0.06 and 96.2%, respectively, under pulsed operation. These metrics confirm that the system achieves the fidelity necessary for the linear computations performed by the optical neuron.
• Figure 4: Demonstration of an all-optical neuron by optically cascading the PPLN and MZI chips for single-hidden-layer ONN learning.a Pulsed pump at phase-matched FH wavelength is launched to an intensity modulator (IM), where the weighted neuron input values are encoded using an arbitrary waveform generator operating at 1 kHz. The modulated light is subsequently coupled into a PPLN chip to realize $\chi^{(2)}$-based nonlinear activation. The nonlinearly transformed pulse is then routed to a silicon MZI chip, which performs the weighted multiplication, and the on-chip inference output is detected at the on-chip photodetection readout. b Decision boundaries generated by the SHG-activated ONN for three binary classification datasets: Moon, Circle, and Gaussian. Pink and green dots represent two separable sets. c Training loss and accuracy curves over 10000 epochs in the differentiable physics-aware digital twin of the ONN. d ROC curves and AUC scores of the ONN inference demonstrate high classification performance: AUC = 0.99 (Moon), 0.95 (Circle), and 0.96 (Gaussian), significantly outperforming the chance level (dashed line). e Schematic diagram of Iris flower dataset consisting of three species. f Two-dimensional map of ONN’s decision surface around training data points, with axes representing directions in feature space based on principal component analysis. g Partial dependence shows predicted probability for a class change along varied feature value. Black vertical line represents the current feature value of the selected instance.
• Figure 5: Multi-layer ONNs with measured $\chi^{(2)}$-based activations in hybrid electro-optical neural computing.a Inference pipeline of the proposed optical neurons on the MedMNIST dataset. The input image is first flattened into a vector and encoded into the amplitudes and phases of optical signals. These signals undergo a linear weighted-sum operation via MZI-based interference, producing hidden-layer representations. The outputs are then nonlinearly activated using the fitted FH response curve derived from experimentally measured pump-depleted SHG data. After passing through multiple layers of MZI–SHG-based transformations, the final output is fed into a softmax layer to generate the prediction distribution. b Example prediction probability distribution for a nevi class generated by the ONN. c Training loss, accuracies, and test accuracy over epochs on the DermaMNIST-C dataset. d The confusion matrix of the MLP model trained on the dataset. e A schematic representation of the Airfoil Self-Noise dataset with five aerodynamic features. f Training loss and test $R^2$ score over 5,000 epochs, demonstrating the ONN's stable convergence. g Residuals plotted against predicted values, with an accompanying histogram showing their distribution centered around zero. The inset displays predicted versus true sound pressure levels, demonstrating strong agreement along the ideal diagonal, indicating accurate model fitting.