Reconfigurable all-optical inference via tunable second-harmonic generation and spin-orbit coupling cascade

Technical Summary: Reconfigurable All-Optical Inference via Tunable Second-Harmonic Generation and Spin-Orbit Coupling Cascade

Problem Statement
Conventional approaches to generating orbital angular momentum (OAM) via optical spin-orbit coupling (SOC) typically rely on rigid architectures, such as tightly focused beams, anisotropic media, or metasurfaces. These methods suffer from a lack of flexibility and tunability, often yielding finite-dimensional spin-OAM superpositions that preclude dynamic mode reconfigurability. Furthermore, existing nonlinear schemes, such as second-harmonic generation (SHG) combined with SOC, largely depend on purely circularly polarized inputs. This constraint confines polarization evolution to singular spin paths, resulting in sparse, discrete OAM spectra and failing to exploit the full potential of nonlinear optics for constructing high-dimensional feature spaces. The challenge remains to generate a continuously tunable, high-dimensional OAM spectrum using simple polarization inputs to support advanced optical computing tasks.

Methodology
The authors propose and experimentally demonstrate a continuously tunable cascaded SOC-SHG mechanism within a single, spatially movable nonlinear crystal (a z-cut barium metaborate, BBO, crystal). The system operates under linearly polarized excitation propagating along the crystal's optical axis.

• Physical Mechanism: The interaction involves a synergistic evolution where the fundamental frequency (FF) field undergoes both degenerate and nondegenerate SHG processes simultaneously with SOC. The FF field excites second-harmonic frequency (SF) fields, where the spin-OAM modes are mapped into new OAM modes governed by nonlinear optical selection rules and crystal symmetries.
• Tunability: The core innovation is the spatial translation of the nonlinear crystal along the optical axis. By varying the focal position ($d_F$) of the incident FF beam relative to the crystal, the authors deterministically control the amplitude distribution of the resulting spin-OAM superposition states.
• Algorithmic Analogy: The physical evolution of the SHG-SOC cascade is mathematically analogous to the kernel function in a Support Vector Machine (SVM). The nonlinear crystal acts as a physical kernel that projects low-order, linearly inseparable input data into a high-dimensional OAM feature space where the data becomes linearly separable.
• Experimental Setup: High-Order Poincaré (HOP) beams or linearly polarized beams are prepared using wave plates and focused into the BBO crystal. The generated SF fields are analyzed using polarization analyzers, vortex phase plates, and single-mode fibers connected to power meters to resolve the spin-OAM spectrum.

Key Contributions

1. Continuous Tunability: The work introduces a method to continuously reconfigure the OAM spectral distribution simply by translating the nonlinear crystal, overcoming the rigidity of previous fixed-architecture SOC platforms.
2. High-Dimensional Feature Expansion: The system expands the input OAM spectrum from a low-dimensional state (e.g., 2D or 4D) into a high-dimensional space (e.g., 12D or higher) containing multiple orthogonal modes (including states like $\pm 4, \pm 6, \pm 2, 0$). This expansion is achieved through the concurrent engagement of degenerate SHG, nondegenerate SHG, and SOC processes.
3. All-Optical Inference Platform: The authors construct a reconfigurable all-optical inference engine where the physical evolution of the light field performs high-dimensional feature extraction, and the crystal position serves as a tunable parameter to modulate the optical kernel function.

Results

• Spectral Evolution: Experimental and theoretical results confirm that the spin-OAM spectrum of the SF field is highly dependent on the focal position ($d_F$).
  • At the input face ($d_F=0$), the spectrum is dominated by direct SHG states ($\pm 4$) and SOC-SHG states ($\pm 6$).
  • At the midplane ($d_F=d/2$), a complex interplay generates a broad spectrum including $\pm 4, 0, \pm 2, \pm 6$.
  • At the output face ($d_F=d$), the spectrum shifts to be dominated by SOC-preceded SHG states ($\pm 0$), with other components vanishing or attenuating.
• Classification Tasks:
  • Iris Dataset: The system successfully classified Iris flowers (Setosa, Versicolor, Virginica). Four-dimensional input data (petal length, petal width, sepal width, sepal length) were encoded into the weights of specific spin-OAM components. By optimizing the focal position to $d_F = 5.000$ mm (minimizing cross-entropy), the system mapped the data into a high-dimensional space where the classes formed distinct, linearly separable clusters detectable via fundamental Gaussian modes.
  • Palmer Penguins Dataset: The architecture was reconfigured for the Palmer Penguins dataset (Adelie, Gentoo, Chinstrap) without hardware changes. Six-dimensional input vectors were encoded, and the optimal focal position was determined to be $d_F = 4.550$ mm. The system successfully categorized the three penguin species based on the resulting optical power distributions.

Significance and Claims
The paper claims to establish a scalable, physically reconfigurable architecture for high-dimensional all-optical computing and neuromorphic photonics. Key advantages highlighted include:

• Hardware Efficiency: Unlike approaches requiring bulky spatial light modulators or fixed holographic masks, this platform achieves reconfigurability through a single passive mechanical degree of freedom (crystal translation).
• Intrinsic Nonlinearity: The cascaded SOC process inherently implements the activation function and feature expansion, eliminating the need for separate activation functions or digital post-processing, thus offering a compact, low-power alternative to electronic systems.
• Scalability: The physical mapping matrix can be scaled by expanding the basis of incident spin-OAM states, allowing the architecture to adapt to higher-dimensional tasks with minimal additional complexity.

The authors conclude that this work deepens the understanding of spin-orbit interactions in nonlinear regimes and paves the way for all-optical inference engines capable of adaptive, high-speed information processing based on intrinsic physical evolution.