Scalable native signed optical computing enabled by dual-wavelength incoherent multiplexing

Technical Summary: Scalable Native Signed Optical Computing via Dual-Wavelength Incoherent Multiplexing

Problem Statement
While incoherent photonic neural networks (PNNs) offer robustness and scalability advantages over coherent architectures by eliminating strict phase-stabilization requirements, they face a fundamental limitation: optical intensity is intrinsically non-negative. Consequently, most existing incoherent multiply-accumulate (MAC) systems are restricted to non-negative signal representations. Supporting signed (bipolar) inputs and weights—a necessity for efficient optimization and convergence in modern machine learning—has previously required hardware-intensive solutions such as spatially separated differential encoding, temporal multiplexing, or complex nonlinear transfer functions. These approaches typically incur hardware overhead that scales with system size or require additional computing cycles, thereby limiting integration density and scalability for large-scale tensor operations.

Methodology
The authors propose a scalable, dual-wavelength incoherent computing architecture implemented on a thin-film lithium niobate (TFLN) platform. The core innovation is a native signed computation scheme that utilizes two spectrally separated optical carriers ($\lambda_1$ and $\lambda_2$) to encode complementary signal components.

• Encoding Mechanism: A bipolar input $X_i \in [-1, 1]$ is decomposed into two non-negative components, $x_i^+$ and $x_i^-$, such that $X_i = x_i^+ - x_i^-$. These components are encoded onto two distinct wavelengths via high-speed electro-optic modulators.
• Shared Path Computation: The two wavelength channels are routed through a shared physical path containing an unbalanced Mach–Zehnder interferometer (MZI) acting as a wavelength multiplexer. This directs both wavelengths to a shared weighting unit (another unbalanced MZI).
• Complementary Weighting: The weighting MZI is designed such that the transmission weights for the two wavelengths, $w_i^+$ and $w_i^-$, exhibit a complementary response ($w_i^+ + w_i^- = 1$).
• Differential Detection: By taking the difference of the optical intensities at the complementary output ports of the weighting MZI, the system computes $(w_i^+ - w_i^-)(x_i^+ - x_i^-) = W_i X_i$. This realizes true four-quadrant optical multiplication within a single incoherent intensity-computing framework.
• Calibration: The system employs a calibration procedure to identify complementary wavelengths ($\lambda_1 = 1550$ nm and $\lambda_2 = 1548.4$ nm) and uses low-frequency triangular-wave modulation to balance the transmission intensities, minimizing offset errors during differential detection.

Key Contributions

1. Native Signed Architecture: The work demonstrates a method to natively support both signed inputs and signed weights without duplicating weighting units or requiring coherent phase stabilization.
2. Constant Hardware Overhead: Unlike previous methods where hardware complexity scales with matrix size, this spectral multiplexing strategy ensures that the additional hardware overhead for signed computation remains constant per multiply–accumulate operation, regardless of the system dimension.
3. TFLN Implementation: The architecture is realized on a TFLN platform using photolithography-assisted chemo-mechanical etching (PLACE), achieving low propagation loss (0.03 dB/cm) and high manufacturing uniformity.

Experimental Results

• Device Performance: The fabricated modulators exhibit a 3-dB electro-optic bandwidth exceeding 40 GHz with a $V_\pi \cdot L$ of 2.25 V·cm. The weighting MZI achieves extinction ratios exceeding 30 dB and demonstrates the required complementary modulation behavior between the two wavelength channels.
• Multiplication Fidelity: Dynamic four-quadrant optical multiplication was tested with random bipolar inputs and weights. The system achieved a standard deviation error of 1.27% for the multiplication output, with no significant error accumulation observed across the full four-quadrant range.
• System-Level Validation:
  • Moons Dataset: The photonic system achieved a classification accuracy of 95.1% on the nonlinear "Moons" dataset, closely matching the 95.4% accuracy of an ideal electronic baseline.
  • MNIST Dataset: For handwritten digit recognition, the system achieved 91.63% accuracy, approaching the 92.81% accuracy of the ideal electronic model.
  • Training Dynamics: Numerical simulations and experimental data confirmed that bipolar (zero-centered) inputs significantly improve training convergence and final accuracy compared to non-negative constraints.

Significance and Claims
The paper establishes a practical route toward scalable incoherent photonic computing systems capable of native bipolar processing. By eliminating the need for duplicated hardware paths or complex temporal encoding for signed operations, the proposed architecture addresses a critical bottleneck in the scalability of optical neural networks. The authors claim that this approach, combined with the ultrahigh bandwidth of the TFLN platform, provides a compact and robust hardware foundation for future large-scale photonic neural networks. The work suggests that this spectral encoding strategy allows for further scaling through hybrid spectral and spatial multiplexing without fundamentally altering the system architecture or increasing the per-channel hardware cost for signed computation.