Optical next generation reservoir computing
Optical NGRC uses light scattering with delayed inputs to implicitly generate polynomial features, enhancing chaotic time series prediction.
Key Findings
Methodology
This work introduces an optical NGRC framework employing light scattering media to realize implicit polynomial feature generation via delayed inputs. The core algorithm encodes data onto a phase modulator, uses scattering matrices for linear mixing, and applies squared-law detection for nonlinear mapping. The system constructs reservoir features without explicit nodes, leveraging the transmission matrix M_s to connect optical features with digital polynomial representations. Training involves linear regression of the readout layer Wout to predict chaotic systems like Lorenz63 and Kuramoto-Sivashinsky, demonstrating short-term accuracy and long-term climate replication. The approach combines physical nonlinearity with computational efficiency, avoiding explicit reservoir construction while maintaining interpretability.
Key Results
- In Lorenz63 forecasting, the optical NGRC achieved a normalized RMSE of 0.0971 with only 4000 training steps, outperforming previous optical RC methods by reducing training data by over 90%. Long-term simulations successfully reproduced attractor structures, confirming the system’s ability to capture statistical properties.
- For Kuramoto-Sivashinsky chaos, the system predicted up to 4 Lyapunov times with an RMSE of 0.2988, using only 6000 training steps, significantly shorter than traditional methods requiring tens of thousands of steps. Power spectral analysis showed high fidelity in long-term behavior.
- As a reservoir observer, the system accurately inferred unmeasured variables from partial data, outperforming spline interpolation, demonstrating its potential in real-world system monitoring and control.
Significance
This optical NGRC paradigm advances physical reservoir computing by leveraging light scattering for high-dimensional, implicit feature generation. It addresses the bottleneck of hardware complexity and training cost in traditional RC, enabling efficient, interpretable, and scalable chaotic system modeling. Its ability to perform accurate short-term prediction and long-term climate reproduction opens new avenues for real-time complex system analysis, with broad implications for physics, engineering, and data science.
Technical Contribution
The key innovation lies in mapping digital polynomial features into a physical optical system via phase modulation and scattering matrices, creating an implicit nonlinear feature space without explicit reservoir nodes. The introduction of the system matrix M_s links optical features to digital representations, enabling efficient training and interpretability. This approach significantly reduces hardware complexity and training data requirements, offering a new framework for scalable, high-dimensional physical reservoir computing.
Novelty
This is the first demonstration of using light scattering media to implement an implicit polynomial feature generator for NGRC, bridging digital and physical reservoir computing. Unlike prior optical RC relying on explicit spatial or delay-based reservoirs, this method exploits optical nonlinearity and delayed inputs to generate rich features efficiently, marking a significant leap in physical neural network design.
Limitations
- The system's performance is sensitive to optical noise and device quantization, which can limit prediction accuracy, especially in noisy environments. Although optimized, high noise levels still pose challenges.
- Current implementation is semi-automatic, requiring digital control for phase modulation and detection; a fully optical system remains a future goal.
- Scaling to higher-dimensional or more complex systems may require larger optical matrices and more sophisticated hardware, increasing complexity and cost.
Future Work
Future efforts will focus on developing fully integrated, all-optical NGRC systems to eliminate digital bottlenecks. Extending the framework to multi-modal data and deep architectures could further enhance expressivity. Applications in climate modeling, financial forecasting, and real-time system monitoring are promising directions, alongside efforts to improve robustness against noise and hardware imperfections.
AI Executive Summary
This study introduces an innovative optical NGRC framework leveraging light scattering media to perform high-dimensional, implicit polynomial feature generation driven by delayed inputs. Unlike traditional reservoir computing, which relies on explicit nonlinear nodes, this approach encodes data onto a phase modulator, passes it through a scattering medium, and detects the resulting speckle intensities. The core idea is to utilize the optical transmission matrix M_s to map optical features into a digital polynomial feature space, enabling efficient, interpretable prediction of chaotic systems.
Experimental results demonstrate that the optical NGRC accurately forecasts Lorenz63 and Kuramoto-Sivashinsky chaotic time series, achieving a normalized RMSE of 0.0971 and 0.2988 respectively, with significantly reduced training data and system size compared to prior methods. Long-term simulations successfully reproduce attractor structures and system 'climate,' confirming the model's ability to capture statistical properties beyond short-term prediction. Moreover, as a reservoir observer, it infers unmeasured variables from partial data, outperforming classical interpolation techniques.
This approach offers a scalable, energy-efficient alternative to conventional optical RC, with broad implications for real-time complex system modeling, physics-inspired machine learning, and optical information processing. Its physical interpretability, combined with high performance, paves the way for future all-optical implementations and advanced architectures, promising transformative impacts across scientific and industrial domains.
Deep Analysis
Background
Complex dynamical systems在科学和工程中扮演重要角色，传统数值模拟和机器学习虽能建模，但在效率和可扩展性方面存在瓶颈。储层计算（RC）作为一种非线性动态系统，通过随机连接和线性回归，已在时间序列预测、分类等任务中展现出优势。光学实现利用光的高速和能效优势，主要包括延迟式和空间散射RC。近年来，数字NGRC通过构造多阶多项式特征提升表达力，但缺乏物理实现方案。本文旨在利用光散射介质实现隐式多阶特征，突破硬件限制，推动光学非线性动力学模拟。
Core Problem
现有光学RC多依赖空间散射或延迟反馈，硬件复杂、难以扩展，且多阶特征的显式构建成本高。数字NGRC虽提升表达力，但缺乏物理路径，限制实际应用。如何在物理系统中高效、隐式生成多阶特征，成为核心难题。解决方案需兼顾性能、可解释性和硬件简洁性，推动光学RC向更复杂、更实用方向发展。
Innovation
提出利用光散射介质实现隐式多阶多项式特征的光学NGRC，结合延迟输入机制，避免显式存储节点。引入系统-给定矩阵M_s，将光学特征映射到数字特征空间，增强模型可解释性。创新在于利用光学非线性映射实现高阶特征生成，简化硬件结构，提升预测能力，减少训练数据和超参数，突破传统光学RC的局限。
Methodology
- �� 输入数据（ut, ut−1）通过相位调制在空间光调制器（SLM）上编码。
- �� 调制光束照射散射介质，利用其传输矩阵实现随机线性混合。
- �� 通过平方律检测获得光强，隐式生成非线性特征。
- �� 采用系统-给定矩阵M_s，将光学特征映射到数字特征空间。
- �� 利用线性回归训练读出层Wout，完成时间序列预测。
- �� 实验中，驱动Lorenz63和Kuramoto-Sivashinsky系统，验证短期预测和长时“气候”复制能力。
Experiments
采用Lorenz63和Kuramoto-Sivashinsky数据集，训练样本分别为4000和6000步。评估指标为NRMSE，误差低至0.0971和0.2988。对比传统光学RC，训练时间和数据量显著减少。长时模拟验证系统“气候”保持。作为观测器，能从部分变量推断未测变量，验证其在复杂系统监测中的潜力。实验环境稳定，设备量化和噪声影响有限。
Results
光学NGRC在Lorenz63和KS系统中实现了超越传统方法的预测性能，短期误差低至0.0971和0.2988，长时模拟成功复制吸引子“气候”。训练数据和时间显著减少，系统规模小，表现优越。作为观测器，准确推断未测变量，验证其在复杂系统监测中的潜力。整体表现显示光学NGRC在高效、可解释、可扩展方面具有巨大优势。
Applications
可应用于气候模拟、金融时间序列、复杂系统监控等领域，适合需要高速、低能耗预测的场景。未来结合深度架构，可实现更复杂任务的实时处理。系统对硬件要求低，易于集成，推动光学信息处理在智能制造、环境监测等行业的应用。
Limitations & Outlook
受光学噪声和设备量化影响，预测精度有限。全光学化仍待实现，当前依赖数字控制。模型在高维复杂系统中表现有限，需扩展特征阶数和系统规模。未来需优化光学器件，提升鲁棒性和扩展性。
Plain Language Accessible to non-experts
想象你在一个工厂里，工厂里有许多机器（代表神经网络中的节点），它们通过传送带（光线）连接。每次工厂接收原料（输入信息），机器们会根据原料的不同，做出不同的反应（非线性处理）。这些反应会经过特殊的滤镜（散射介质），让信息变得更加丰富和复杂。最后，工厂的管理者（读出层）会根据机器的反应，决定下一步的行动。这个工厂不用自己建很多机器，而是用光线和散射介质来实现复杂的计算。这样既快又省能，还能处理非常复杂的问题，比如天气预测、金融分析等。
ELI14 Explained like you're 14
想象你在玩一个超级复杂的拼图游戏，你有一堆碎片（数据），但拼图的规则很难理解。这个光学系统就像一个神奇的魔法盒子，把碎片放进去，然后用光线照亮它。光线经过这个盒子后，会变成一堆奇怪的光斑（散射的光），这些光斑其实藏着很多拼图的秘密。你不用自己拼图，只需要用特殊的镜头（读出层）观察这些光斑，然后告诉你拼图的样子。这个方法比传统的拼图更快、更聪明，还能拼出以前拼不出来的图。它可以帮我们预测天气、金融市场，甚至理解复杂的自然规律，就像魔法一样厉害！
Abstract
Artificial neural networks with internal dynamics exhibit remarkable capability in processing information. Reservoir computing (RC) is a canonical example that features rich computing expressivity and compatibility with physical implementations for enhanced efficiency. Recently, a new RC paradigm known as next generation reservoir computing (NGRC) further improves expressivity but compromises its physical openness, posing challenges for realizations in physical systems. Here we demonstrate optical NGRC with computations performed by light scattering through disordered media. In contrast to conventional optical RC implementations, we drive our optical reservoir directly with time-delayed inputs. Much like digital NGRC that relies on polynomial features of delayed inputs, our optical reservoir also implicitly generates these polynomial features for desired functionalities. By leveraging the domain knowledge of the reservoir inputs, we show that the optical NGRC not only predicts the short-term dynamics of the low-dimensional Lorenz63 and large-scale Kuramoto-Sivashinsky chaotic time series, but also replicates their long-term ergodic properties. Optical NGRC shows superiority in shorter training length, increased interpretability and fewer hyperparameters compared to conventional optical RC based on scattering media, while achieving better forecasting performance. Our optical NGRC framework may inspire the realization of NGRC in other physical RC systems, new applications beyond time-series processing, and the development of deep and parallel architectures broadly.