🔬 optics

Artificial Intelligence Reshapes Microwave Photonics

This review paper provides the first comprehensive overview of how artificial intelligence is fundamentally reshaping microwave photonics by revolutionizing the design, simulation, fabrication, testing, deployment, and operation of systems across signal generation, transmission, processing, and detection.

Original authors: Peng Li, Xihua Zou, Jia Ye, Wei Pan, Lianshan Yan

Published 2026-05-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Peng Li, Xihua Zou, Jia Ye, Wei Pan, Lianshan Yan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: From Bulky Boxes to Smart Chips

Imagine Microwave Photonics (MWP) as a super-fast delivery service. Instead of using slow, heavy trucks (traditional electronics) to carry data, it uses light beams (photons) to zip information around at incredible speeds. This is great for things like 5G, radar, and sensing, but for a long time, the "trucks" were huge, clunky, and required a team of mechanics to keep them running smoothly.

The paper describes how this field is evolving through three "eras":

MWP 1.0 (The "Bulky" Era): Like a massive warehouse full of separate, disconnected machines. It worked, but it took up a whole room and needed constant manual tweaking.
MWP 2.0 (The "Integrated" Era): Like shrinking that warehouse down to a single, sleek smartphone chip. Everything is smaller and more stable, but it's still a bit rigid. If the environment changes (like temperature or signal noise), the chip gets confused and needs a human to fix it.
MWP 3.0 (The "Intelligent" Era): This is the new era the paper focuses on. It's like giving that smartphone chip a brain. By adding Artificial Intelligence (AI), the system can now "think," learn from its mistakes, and fix itself in real-time without a human mechanic.

How AI is Changing the Game

The paper breaks down exactly how this "brain" helps in three main areas:

1. Creating Signals: The Improvisational Jazz Musician

In the past, making specific microwave signals was like trying to play a complex song on a piano where every key was stuck unless you manually adjusted the tension of the strings.

The AI Solution: Now, AI acts like a jazz musician who can instantly improvise. Instead of following a rigid script, the AI learns the "rules of physics" from data. It can instantly generate complex, chaotic signals needed for radar or secure communication, adjusting the "notes" (frequencies) in milliseconds to fit the situation perfectly.

2. Sending Signals: The Self-Driving Car

Sending data through fiber optic cables is like driving a car on a road that changes its shape every second due to weather (heat, bends, noise).

The Problem: Traditional systems are like drivers who memorize a map. If the road changes, they crash or get lost because they don't know how to react.
The AI Solution: AI acts like a self-driving car. It doesn't just follow a map; it uses cameras (data) to see the road changing in real-time. It automatically steers around potholes (nonlinear distortions) and adjusts its speed to keep the ride smooth. The paper shows that AI can predict how signals will behave and fix errors before they even happen, making the connection much faster and more reliable.

3. Processing and Detecting Signals: The Master Detective

When the signal arrives, it often looks like a scrambled puzzle. Traditional systems use a fixed set of rules to unscramble it, which is slow and often fails if the puzzle is too messy.

The AI Solution: AI is like a master detective. It doesn't just look for one clue; it learns from thousands of past cases to recognize patterns instantly.
- For Radar: It can look at a messy echo and instantly say, "That's a bird, not a plane," with much higher accuracy than old systems.
- For Interference: Imagine trying to hear a whisper in a noisy room. AI can act like a noise-canceling headphone that learns exactly what the noise sounds like and subtracts it, leaving only the whisper clear.
- For Sensing: It can detect tiny changes in temperature or humidity, even if the sensor itself is drifting or acting up, by "knowing" what the data should look like.

The Future: The "AI Co-Scientist"

The paper concludes by looking ahead to what this means for the future. It suggests we are moving toward a world where:

AI Co-Scientists: Instead of humans spending weeks designing and testing a new device, an AI agent could do the work. It would run thousands of virtual experiments, figure out the best design, and even build it, acting as a tireless assistant to human engineers.
Self-Healing Systems: The devices themselves will become "smart." They will be able to calibrate themselves, fix their own errors, and adapt to new environments without anyone touching them.

Summary

In short, this paper argues that Microwave Photonics is no longer just about building faster hardware. By combining it with Artificial Intelligence, we are creating systems that are not only faster and smaller but also autonomous. They can learn, adapt, and optimize themselves, turning a once-clunky, manual technology into a smart, self-driving network for the future.

Technical Summary: Artificial Intelligence Reshapes Microwave Photonics

Problem Statement
Microwave Photonics (MWP) has evolved from a bandwidth-driven era (MWP 1.0) to an integrated era (MWP 2.0), characterized by photonic integrated circuits (PICs). However, as integration density and functional complexity increase, traditional MWP systems face significant bottlenecks. These include:

Optimization Challenges: Large-scale monolithic integration introduces intricate interdependencies among parameters (lasers, modulators, filters, detectors), making adaptive and real-time optimization difficult in dynamic environments.
Modeling Limitations: Conventional signal generation and transmission rely on deterministic physical models (e.g., Lang–Kobayashi equations, split-step Fourier methods) that are computationally intensive, lack flexibility for varying signal formats, and struggle with complex nonlinear distortions.
Calibration and Stability: Traditional systems often require precise manual calibration, suffer from environmental sensitivity (e.g., temperature drift, phase noise), and exhibit long-term performance instability.
Processing Constraints: Fixed architectures and complicated calibration procedures limit the efficiency of signal processing, detection, and interference cancellation.

Methodology
The paper proposes a paradigm shift to "MWP 3.0," an intelligent era where Artificial Intelligence (AI), specifically Machine Learning (ML) and Deep Learning (DL), is deeply integrated into MWP systems. The methodology involves replacing or augmenting traditional analytical models and manual control loops with data-driven approaches:

Data-Driven Modeling: Utilizing neural networks (e.g., Feedforward Neural Networks, Transformers, LSTM) to learn high-order implicit relationships between inputs and outputs directly from experimental or simulated data, bypassing the need for explicit physical equations.
End-to-End (E2E) Optimization: Treating the entire communication link (transmitter, channel, receiver) as a unified system trained via backpropagation to find globally optimal configurations.
Reinforcement Learning (RL): Employing RL agents to autonomously adjust system parameters (e.g., variable optical attenuators, delay lines) in real-time based on feedback metrics like Bit Error Rate (BER) or Error Vector Magnitude (EVM).
Inverse Design: Using DL architectures (e.g., ResNet) to solve inverse problems, such as determining the required seed temporal profile to achieve a target optical frequency comb spectrum.

Key Contributions and Results
The review systematically categorizes AI-enabled advancements across three domains:

1. AI for MWP Signal Generation

Chaotic Signal Generation: A Nonlinear Optoelectronic Oscillator (NOEO) combined with AI generates four parallel, orthogonal chaotic sequences (18.18 GHz bandwidth) for all-photonic reinforcement learning accelerators.
Semiconductor Laser Control: A Feedforward Neural Network (FNN) emulates the Lang–Kobayashi equations to predict Period-One (P1) oscillation frequencies with sub-megahertz accuracy. This allows for rapid determination of optimal control parameters (milliseconds on CPU, nanoseconds on FPGA), improving efficiency by over three orders of magnitude compared to numerical solvers.
Optical Frequency Combs: A 1D-ResNet architecture enables the inverse mapping of target comb spectra to seed temporal profiles, achieving programmable generation of diverse spectral envelopes (Gaussian, parabolic, etc.) with errors of $10^{-1}$ to $10^{-2}$ .

2. AI for MWP Transmission

Link Modeling: Deep Neural Networks (DNNs) with multi-head attention mechanisms and Transformer-based models predict optical signal propagation with high fidelity (e.g., NMSE = 0.000819 for 2-GBaud 16QAM), outperforming traditional sequential processing and offering superior generalization across distances and modulation formats.
Nonlinear Equalization:
- LSTM-based Detection: A Deep Learning Detection with Reference Tone (DLD-RT) system reduces BER from $10^{-1}$ to $10^{-5}$ in unlocked heterodyne links, outperforming MLP and CNN architectures in phase noise robustness.
- FTnet: A Fourier Layer Transformer network performs end-to-end demodulation, achieving a 3–6 dB improvement in receiver sensitivity and BER below the 7% FEC limit.
System Optimization: End-to-end frameworks using Transfer Learning and Self-Supervised Learning (SSL) jointly optimize transmitters and receivers. These systems demonstrated 3.5 dB (single-user) and 1.5 dB (three-user) improvements in receiver sensitivity for millimeter-wave-over-fiber links.

3. AI for MWP Signal Processing & Detection

Photonic ADC: Integrating DL correction networks with photonic front-ends improved the Effective Number of Bits (ENOB) from 4.6 to 7.3 (and up to 9.24 with ultra-low-jitter lasers) and achieved a Spurious-Free Dynamic Range (SFDR) of 71 dB.
Target Recognition: Combining a 10 GHz photonic ADC with a Convolutional Neural Network (CNN) for radar echo analysis achieved 93.05% target recognition accuracy, approximately 25% higher than traditional electronic systems.
Measurement:
- IFM: Stacking ensemble learning (Random Forest, SVM, etc.) enabled frequency measurement from 1–40 GHz with average errors below 5 MHz.
- AOA: An LSTM-DNN system achieved a Mean Absolute Error (MAE) of 0.1438° for angle-of-arrival estimation over a $\pm80^\circ$ range, eliminating complex phase calibration.
Sensing & Cognition: Multi-agent RL algorithms successfully coordinated frequency switching in high-speed railway RoF links to avoid electromagnetic interference. ML models (SVR, CNN, NTK) enhanced multi-parameter sensing (temperature/humidity) accuracy by factors of 3 to 9.
Interference Cancellation: RL agents automatically adjusted optical components to suppress self-interference. A single-path system achieved 20.18 dB suppression over 800 MHz in eight steps; a multipath DDPG system achieved 24 dB suppression over 1 GHz in five steps.

Significance and Claims
The paper claims that the integration of AI and MWP marks the transition to MWP 3.0, an era defined by the convergence of AI's autonomous operation, PIC miniaturization, and ultra-large bandwidth.

Transformation of Workflow: The authors assert that AI is revolutionizing the design, simulation, fabrication, testing, deployment, and maintenance of MWP systems, moving them from static, manually tuned architectures to adaptive, data-driven, and self-optimizing platforms.
Performance Gains: The review highlights that AI enables exceptional efficiency, autonomous operation, and performance beyond traditional systems, specifically in handling nonlinearities, dynamic environments, and complex optimization tasks.
Future Outlook: The paper posits that the future lies in AI co-scientists/assistants for closed-loop system optimization and AI-embedded MWP systems (neuromorphic photonics) that possess self-calibrating and self-optimizing capabilities directly on-chip. It emphasizes the need for multimodal AI agents to bridge the gap between photonics, electronics, and other physical domains to achieve full-spectrum intelligence.

The review serves as the first comprehensive overview of AI-enabled MWP, providing a systematic summary of state-of-the-art advances and offering insights for both the academic community and the broader public.