Santosh Kumar

We present a loss-swept calibration procedure for parallel superconducting nanowire single-photon detectors (P-SNSPDs) using weak coherent states under controlled optical attenuation. An O-band (∼1300 nm) coherent state from a 20 MHz mode-locked laser is directed through a voltage-controlled variable optical attenuator (VOA), acting as a characterized amplitude-damping channel. A Fisher information analysis of the N-element multiplexed detector model shows that the mean photon number μ ₀ and the detection efficiency η _det are not individually identifiable from click statistics alone: the Fisher correlation is ρ ( μ ₀ , η _det ) = 1.000. We therefore adopt the effective detection rate λ ₀ ≡ μ ₀ η _det as the primary calibration parameter, obtaining λ ^ ₀ =0.3824 across 31 voltage settings; η _det carries zero Fisher information from click statistics and is anchored by an independent power-meter measurement of optical power through a calibrated cascaded-attenuator chain of 86.0 dB, giving μ ^ ₀ =0.654 photons/pulse at the detector and η ^ _det = λ ^ ₀ / μ ^ ₀ =0.585. We further extend the model to account for a flux-dependent detection efficiency η ( μ ) using a saturating dead-time model, and introduce a Fisher-orthogonal anchor parametrization ( η _L , η _H ) that separates the efficiency parameters across the attenuation sweep. The anchor efficiencies are well determined ( η ^ _L =0.9145, η ^ _H =0.9435), but the saturation scale they imply is not: with the per-voltage mean photon numbers left free, as the anchor fit requires, μ _sat is degenerate with them and is not identifiable from click statistics at any precision. Holding the flux instead at its calibrated value, a profile likelihood bounds the saturation scale only as μ _sat ≳0.93 at 95% confidence, the best fit being the no-saturation limit. We therefore report a bound rather than a value, and identify the limited dynamic range of the sweep ( μ ^ _L / μ ^ _H ≈3.4), together with the external anchoring of the per-voltage flux, as the two requirements for determining μ _sat in future work. Deviations of λ ( V ) from the linear VOA model show small systematic signatures whose sign and magnitude are consistent with first-order pixel-coincidence saturation in the N = 4 pixel array. We also quantify band-assignment crosstalk as a systematic on λ ^ ₀ , and find it to be the leading one under the log-scale cost function adopted here.

We present a hybrid computational framework that integrates optical reservoir computing with transformer architectures, trained via a hybrid backpropagation-feedback alignment scheme, for sequence modeling and text generation. By leveraging the transient dynamics of the optical reservoir, temporal signals are preprocessed in the analog domain, enhancing feature saliency and reducing input complexity. Feedback alignment, a biologically inspired alternative to conventional backpropagation, enables hardware- efficient and scalable training. Experimental results on benchmark datasets demonstrate that the proposed hybrid optical transformer achieves competitive performance with stable training dynamics. A key advantage of this approach is its improved energy efficiency as network depth increases, enabled by the recurrent mapping of optical signals onto a modulator. This architecture supports the realization of very deep neural networks using simple diffractive optical systems, where vector–matrix multiplications at each optical layer are executed in a single pass. Although the hybrid model requires more training epochs than a baseline transformer, it achieves optimal perplexity values comparable to electronic baselines while maintaining semantic coherence in generative tasks, thereby validating the feasibility of the hybrid approach for scalable sequence modeling.

We propose and demonstrate a proof-of-concept hybrid photonic–digital neural network that exploits optical frequency comb lines and their second-order nonlinear interactions for both classification and image generation. In this architecture, input features are encoded in the amplitudes of individual comb lines, while their relative phases serve as trainable parameters. For classification, it achieves ∼98% average accuracy on a 40-sample make-moons dataset, with the mean square error reduced from 0.7 to 0.05 post-training. For image generation, a conditional variational autoencoder is implemented using a small dataset of 18 MNIST digits over 3 categories, generating new digits with a reconstruction loss of 0.75. These results establish the feasibility of combining frequency comb lines and nonlinear optics for both discriminative and generative neural network tasks.

Hopfield neural networks, classic examples of associative memories, are well-established models for pattern storage and retrieval. However, the original Hopfield network, characterized by quadratic interactions among neurons, exhibits limited storage capacity. To enhance this capacity, researchers have introduced nonlinear functions—such as polynomial and exponential functions—to generalize the network’s energy landscape. In this work, we propose an optical implementation of the Hopfield network that employs optical parametric amplification to realize a hyperbolic interaction function. We numerically simulate the pattern storage and retrieval dynamics of our optical model and compare its storage capacity with that of Hopfield networks using polynomial interaction functions. Our results show that the storage capacity of the proposed model increases exponentially with the number of neurons and depends on optical parameters such as pump laser power and nonlinear medium properties. This dependence enables further enhancement of storage capacity by tuning the physical hardware alone, without modifying the network architecture.

Per- and polyfluoroalkyl substances (PFAS) pose significant environmental and health concerns, necessitating their efficient and accurate identification to facilitate their eventual mitigation from the environment. Surface-enhanced Raman spectroscopy (SERS) enables highly sensitive and precise molecular identification, but trace-level detection of chemicals and fluorescence interference remain significant challenges. Here, we present a uniform 3D AgNP@Si substrate for SERS, leveraging photon counting to achieve susceptible and low-fluorescence detection. This approach enables the detection of PFAS, including perfluorooctanoic acid (PFOA) and perfluorooctane sulfonate (PFOS), at concentrations as low as 10⁻¹⁵ M, with Rhodamine 6G (R6G) used as a model analyte. Additionally, the quantitative analysis demonstrated a strong logarithmic relationship between Raman intensity and analyte concentration, with high correlation coefficients (R² = 0.98 for R6G and 0.97 for PFOA and PFOS). This pioneering approach offers a promising alternative to current analytical techniques for monitoring PFAS and other contaminants in the environment.

Modern Hopfield Neural Networks (HNNs), also known as Dense Associative Memories (DAMs), enhance the performance of simple recurrent neural networks by leveraging the nonlinearities in their energy functions. They have broad applications in combinatorial optimization, high-capacity memory storage, deep learning transformers, and correlated pattern recognition. Thus far, research on DAMs has been primarily theoretical, with implementations limited to CPUs and GPUs. In this work, for the first time to our knowledge, we propose and experimentally demonstrate a nonlinear optical Hopfield neural network (NOHNN) system for realizing DAMs using correlated patterns. Our NOHNN incorporates effective 2-body and 4-body interactions in its energy function. The inclusion of 4-body interaction scores a minimum ten-fold improvement in the number of uncorrelated patterns that can be stored and retrieved, significantly surpassing the traditional capacity limit for traditional HNNs. For correlated patterns, depending on their average correlation, up to 50 times more patterns can be stored compared to traditional HNNs. To test the system's robustness, the benchmark testing is performed on MNIST handwritten digit patterns. The results show a 5.5 times improvement in the pattern storage along with the retrieval of cleaner and less noisy patterns. These results highlight the potential of nonlinear optical DAMs for practical applications in challenging big-data optimization, computer vision and graph network tasks.

Hopfield neural networks (HNNs) promise broad applications in areas such as combinatorial optimization, memory storage, and pattern recognition. Among various implementations, optical HNNs are particularly interesting because they can take advantage of fast optical matrix-vector multiplications. Yet their studies so far have mostly been on the theoretical side, and the effects of optical imperfections and robustness against memory errors remain to be quantified. Here we demonstrate an optical HNN in a simple experimental setup using a spatial light modulator with 100 neurons. It successfully stores and retrieves 13 patterns, which approaches the critical capacity limit of α _c = 0.138. It is robust against random phase flipping errors of the stored patterns, achieving high fidelity in recognizing and storing patterns even when 30% pixels are randomly flipped. Our results highlight the potential of optical HNNs in practical applications such as real-time image processing for autonomous driving, enhanced AI with fast memory retrieval, and other scenarios requiring efficient data processing.

Spatial-mode projective measurements could achieve super-resolution in remote sensing and imaging, yet their performance is usually sensitive to the parameters of the target scenes. We propose and demonstrate a robust classifier of close-by light sources using optimized mode projection via nonlinear optics. Contrary to linear-optics based methods using the first few Hermite–Gaussian (HG) modes for the projection, here the projection modes are optimally tailored by shaping the pump wave to drive the nonlinear-optical process. This minimizes modulation losses and allows high flexibility in designing those modes for robust and efficient measurements. We test this classifier by discriminating one light source and two sources separated well within the Rayleigh limit without prior knowledge of the exact centroid or brightness. Our results show a classification fidelity of over 80% even when the centroid is misaligned by half the source separation, or when one source is four times stronger than the other.