Node-reduction through Joint Optimization of Input and Readout Layers in Photonic Reservoir Equalization

1. Problem Statement

Context: Optical communication systems (specifically Intensity-Modulation/Direct-Detection or IM/DD) face increasing challenges with intersymbol interference (ISI) caused by chromatic dispersion, bandwidth limitations, and device nonlinearities as symbol rates and transmission distances grow.
Limitations of Current Solutions:

• Classical Equalizers: Feed-forward equalizers (FFEs) and Volterra filters can address ISI but suffer from rapidly increasing implementation costs and complexity as the channel memory and nonlinearity increase.
• Photonic Reservoir Computing (RC): While promising due to high bandwidth and low latency, standard photonic RC implementations typically treat the input mapping as fixed, random, or heuristic. Learning is confined solely to the output (readout) layer.
• The Core Issue: To achieve strong performance with fixed-input RC, relatively large network sizes (many nodes) are required. In integrated photonic circuits, increasing the node count directly increases hardware complexity, requiring more interferometric elements, waveguide interconnects, and routing overhead.

Research Question: Can the joint optimization of the input layer (trainable optical weights) and the readout layer significantly improve the node-scaling behavior of a photonic reservoir, thereby allowing for a reduction in the number of physical nodes required to achieve a specific equalization performance?

2. Methodology

Simulation Framework:
The authors employed a hybrid simulation approach to ensure physical accuracy:

• Channel Modeling: VPIphotonics was used to simulate the optical communication link, generating distorted signals with chromatic dispersion, Kerr nonlinearities, and noise.
• Reservoir Modeling: Photontorch (Python library) simulated the all-optical reservoir.
• Signal Parameters: 28 GBd NRZ-OOK transmission, 8 samples per symbol, covering distances up to 200 km.

Architecture:

• Reservoir Core: A fixed, four-port all-optical silicon photonic reservoir. It uses a planar, power-efficient topology where nodes are 3×3 multimode interferometers (MMIs). Internal connections act as delay lines (half the bit period). The internal weight matrix ($R$) is fixed and determined by fabrication (waveguide propagation), making it a "fixed dynamical system."
• Trainable Layers:
  • Readout Layer ($w_{out}$): Standard trainable complex weights combining reservoir states.
  • Input Layer ($w_{in}$): The novel contribution. Instead of fixed/random coupling, the input signal is distributed via programmable complex optical weights (amplitude and phase) implemented by Mach-Zehnder Interferometers (MZIs).
• Optimization: The BFGS algorithm was used to minimize Mean Squared Error (MSE).

Experimental Design & Fairness:
To ensure a fair comparison of "hardware cost," the authors compared two scenarios with an equal total number of trainable weights (degrees of freedom):

1. Joint Optimization: An $N$-node reservoir with both trainable input ($w_{in}$) and readout ($w_{out}$) layers.
2. Readout-Only Baseline: A $2N$-node reservoir with only a trainable readout layer (fixed input).

• Rationale: Both systems use the same total number of weights, but the Joint system uses half the physical reservoir nodes and recurrent links.

Input Policies:
Three strategies for selecting which nodes receive the input signal were tested:

• All: All available input nodes are driven.
• Center: Driven nodes are concentrated near the geometric center of the reservoir.
• Spread: Driven nodes are distributed as broadly as possible across the grid.

3. Key Contributions

1. Joint Input-Readout Optimization: Demonstrated that treating the optical input layer as a trainable design variable (rather than fixed/random) significantly enhances reservoir performance.
2. Node-Scaling Reduction: Proved that a jointly optimized system can achieve the same Bit Error Rate (BER) as a readout-only system with twice the number of nodes, effectively halving the required photonic hardware (nodes and interconnects).
3. Memory Extension: Showed that joint optimization effectively extends the usable memory of the reservoir, leading to over three orders of magnitude improvement in memory-intensive tasks.
4. Hardware Efficiency: Established that the performance gain is not merely due to an increase in parameter count, but due to the optimal allocation of trainable resources (reshaping the feature map) within a smaller physical footprint.

4. Key Results

Performance vs. Transmission Distance:

• At a fixed receiver OSNR of 30 dB, an 8-node jointly optimized reservoir consistently outperformed an 8-node readout-only reservoir by 1–2 orders of magnitude in BER across all tested distances.
• Crucially, the 8-node joint system approached the performance of a 16-node readout-only system (which has the same total weight count) for short-to-mid reach distances (up to ~180 km).
• Beyond 180 km, all systems degraded, but the joint system maintained a clear advantage.

Node Scaling (200 km Scenario):

• At a challenging operating point (200 km, 15 dBm launch power), the jointly optimized systems showed a much steeper improvement in BER as node count increased compared to readout-only variants.
• 16-node Joint System: Achieved a median BER of $3.27 \times 10^{-3}$.
• 16-node Readout-Only System: Achieved a median BER of $1.31 \times 10^{-1}$ (a gap of nearly two orders of magnitude).
• 24-node Joint System: Achieved a median BER of $3.54 \times 10^{-5}$, while the best 24-node readout-only system was at $4.15 \times 10^{-2}$.

Input Policy Impact:

• The "All" policy (driving all input nodes) yielded the best scaling trend for joint optimization, suggesting that access to a broader set of injection channels allows the system to better reshape the reservoir's delayed interference features.
• The "Center" policy performed the worst among joint configurations.

Comparison with Classical Equalizers:

• Starting at 16 nodes, the jointly optimized reservoir achieved performance 1–2 orders of magnitude better than a complexity-matched FFE and a second-order Volterra filter.

5. Significance and Conclusion

Hardware Implications:
The study demonstrates that joint optimization allows for a 2x reduction in the physical size of the reservoir (nodes and recurrent links) while maintaining equal performance. In integrated photonics, where node count dictates the number of interferometers and waveguide routing complexity, this is a critical step toward scalable, low-cost photonic equalizers.

Mechanism of Improvement:
The trainable input layer does not add memory capacity to the reservoir; rather, it reshapes how the fixed reservoir dynamics are excited. By optimizing the injection of the distorted signal, the system exposes a more useful subset of the reservoir's delayed interference features to the readout layer, making the existing hardware more efficient.

Trade-off:
The benefit comes at the cost of increased offline optimization effort (training time) to find the optimal input weights. However, this is a favorable trade-off for reducing the physical footprint and power consumption of the inference-time hardware.

Future Work:
The authors suggest experimental validation, explicit area-control trade-off analysis, and extending the approach to other modulation formats and reservoir architectures as natural next steps.