A Digital Surrogate Framework for End-to-End Optimization of IM/DD Optical Communication Systems

1. Context and motivation High-capacity connectivity is a cornerstone of the ongoing digital transformation of industry and society. Data-center interconnects, campus networks, metro aggregation, and increasingly distributed industrial infrastructures must deliver rising throughput under strict constraints on cost, energy consumption, and operational complexity. In this landscape, intensity modulation and direct detection (IM-DD) remains highly relevant because it enables high data rates with comparatively simple opto-electronic hardware, and it can be scaled to higher symbol rates by combining bandwidth-efficient modulation formats with strong signal processing. This thesis was developed within the doctoral programme in Artificial Intelligence and Big Data for Industry 5.0 and Circular Economy (Universitas Mercatorum, Rome) and co-funded in an industrial context (Almaviva S.p.A.). The overarching motivation is to enable systematic and reproducible optimization of short-reach IM-DD links in a manner consistent with practical devices and laboratory measurement workflows. The core idea is to replace ad-hoc “one-off” simulations and isolated component studies with an end-to-end, device-aware modeling framework that supports analysis, calibration, and data generation for learning-assisted optimization. 2. Problem statement and research gap A persistent methodological gap appears in the IM-DD literature. Many studies adopt system-level models where mixed-signal hardware is idealized or absorbed into simplified aggregated impairments; other studies develop accurate component models but do not integrate them into a modular transmitter–channel–receiver chain intended for end-to-end analysis, optimization, and structured dataset generation. In realistic bandwidth-limited links, limiting behavior emerges from strong coupling among transmitter mixed-signal effects (DAC bandwidth, quantization and static nonlinearities, driver response, modulator transfer nonlinearity), optical propagation effects, optical amplifier dynamics and ASE, and receiver front-end and ADC effects. Treating these elements as separable blocks with independent margins can miss interaction effects that dominate the achievable BER/OSNR operating point and the complexity required by receiver DSP. Machine learning (ML) has been widely explored for equalization, predistortion, and impairment compensation. However, many ML approaches remain effectively black-box: they may improve BER or EVM under a specific training distribution, provide limited interpretability, and degrade under domain shift between training and deployment conditions. This motivates a modeling framework where device behavior is exposed through physically meaningful parameters and where generated data carries structured metadata connecting device settings, waveform features, and performance outcomes, enabling controllable optimization and more robust generalization. 3. Hypothesis, objectives, and research questions The central hypothesis of the thesis is that a modular, device-aware, end-to-end digital surrogate of an IM-DD link—built with disciplined interfaces, strict parameterization, and structured metadata—can preserve the coupled nature of dominant impairments while remaining computationally feasible and calibration-ready, and can thereby support robust analysis and learning-assisted optimization. Four objectives guide the work: (O1) define a rigorous end-to-end architecture with reproducible parameter governance; (O2) model dominant transmitter, channel, and receiver non-idealities as parameterized modular blocks; (O3) enable ML readiness through structured metadata and well-defined observation points; and (O4) demonstrate the approach through receiver DSP analysis and an experimental study on bipolar PAM with learning-based equalization. These objectives are addressed through four research questions: RQ1 asks how to design a mixed-signal end-to-end surrogate that preserves physically meaningful coupling across transmitter, channel, and receiver while remaining modular; RQ2 asks how to represent dominant non-idealities at sample-level fidelity while keeping the surrogate feasible and calibratable; RQ3 asks how to generate structured datasets and metadata that enable interpretable and controllable optimization, including hybrid DSP/ML strategies; and RQ4 asks how to validate the surrogate against experiments, quantify mismatch and uncertainty, and mitigate domain shift when using synthetic or partially calibrated data. 4. Digital surrogate architecture and methodology The thesis introduces an end-to-end surrogate architecture organized as modular processing blocks spanning digital, electrical, and optical domains, implemented as a blockcomposable MATLAB software system. Each block follows a standard input–output contract, and the signal is carried through a canonical waveform container to avoid ambiguity about what is processed at each stage. A dedicated time manager enforces a single authoritative time base and sampling-rate discipline across the chain, performs controlled resampling when required, and supports consistent alignment for visualization and scope-like virtual instruments. The surrogate is parameter-driven: symbol rates, sampling rates, filter bandwidths, noise levels, and nonlinearity settings are defined centrally rather than hard-coded. This enables reproducibility, systematic sweeps, and transparent ablation studies, and keeps device behavior visible as controllable “knobs.” The architecture separates visualization-oriented metadata (for consistent plotting and measurement-like diagnostics) from ML and optimization metadata (for dataset generation and learning experiments), and supports lightweight execution modes for debugging as well as full-fidelity modes for data creation. Internal consistency checks and endto- end regression “golden cases” are used to preserve reliability as the model evolves. 5. Transmitter and channel modeling On the transmitter side, the surrogate includes reproducible bit generation and mapping for PAM formats, including bipolar variants, followed by upsampling and pulse shaping to generate bandwidth-limited waveforms. A central contribution is the device-aware treatment of the DAC chain as a pipeline rather than a single ideal quantizer. The DAC module includes sampling-rate matching at the converter boundary, optional pre-emphasis as a digital conditioning stage, affine remapping into the DAC electrical range with margin control, uniform quantization with explicit coding conventions, optional modeling of static nonlinearity (INL/DNL), zero-order hold (ZOH) reconstruction, thermal and reference-like noise based on physically grounded formulations, and output-network or analog filtering represented as an LTI reconstruction path. Electro-optic modulation is modeled through a parameterized Mach–Zehnder modulator (MZM) transfer function with explicit bias control, enabling analysis of nonlinear distortion and bias sensitivity. In the optical path, attenuation and chromatic dispersion are included in a form consistent with IM-DD analysis, where dispersion can induce power fading after squarelaw detection. Noise can be introduced through controlled abstractions at selected observation points (useful for targeted studies) or through physically tied mechanisms in amplification and detection (useful for device-aware analysis and calibration workflows). 6. Receiver front-end and ADC modeling On the receiver side, the thesis develops a device-aware model of the pre-amplified frontend and the digitization chain, emphasizing mechanisms that introduce memory or reshape noise statistics. Optical amplification is represented via a reduced-order semiconductor optical amplifier (SOA) abstraction capturing small-signal gain, saturation or gain compression, and recovery dynamics, thereby reproducing pattern-dependent behavior and nonlinear memory. ASE is modeled as stochastic optical-field noise parameterized by gain and noise figure (or spontaneous emission parameters), and its interaction with optical filtering and direct detection is handled consistently so that the resulting intensity noise statistics reflect beat-noise mechanisms relevant to pre-amplified IM-DD. The photoreceiver model captures responsivity-based conversion from optical power to photocurrent and includes dominant noise mechanisms. Shot noise is modeled with spectral density proportional to current (including optional dark current), thermal noise is computed using the Johnson–Nyquist formulation and integrated over a noise-equivalent bandwidth, and additional TIA noise densities can be included as input- or output-referred terms. Receiver bandwidth limitation is represented as an LTI response that introduces deterministic ISI and shapes the noise spectrum, making the overall discrete-time channel seen by DSP a coupled outcome of analog responses and sampling. The ADC is treated as a controlled interface. Sampling and resampling operations are explicit, aperture jitter is modeled as a sampling-time perturbation that yields amplitude noise proportional to waveform slope and therefore links timing uncertainty to SNR and EVM degradation, quantization is implemented with explicit conventions (including mid-rise versus mid-tread distinctions), and static ADC nonlinearity (INL/DNL) can be included. Anti-alias filtering and normalization steps are integrated so that digitization effects can be studied jointly with optical and analog impairments. 7. Digital receiver DSP and experimental demonstration Receiver DSP is framed through a discrete-time view where the end-to-end IM-DD link is a finite-memory (often nonlinear) channel. This perspective motivates equalization and robust decision strategies and provides a consistent interface between the surrogate and algorithmic processing. The thesis reviews canonical DSP steps such as alignment and resampling, normalization and scaling, filtering choices, and equalization, and discusses classical equalizers (symbol-rate and fractionally spaced FFE/DFE) together with nonlinear and learning-assisted approaches. A particular focus is bipolar PAM. Under direct detection, square-law detection can obscure sign information; the thesis highlights that fractionally spaced sampling can recover additional information. With two samples per symbol, half-symbol samples contain cross-term contributions that preserve relative sign between adjacent bipolar symbols, enabling bipolar recovery when combined with appropriate DSP or learning-based equalization. This observation provides a concrete link between waveform physics, sampling strategy, and the achievable performance of receiver algorithms. The concept is demonstrated experimentally using an offline MATLAB waveform played by a 100 GSa/s-class DAC, amplified by a broadband driver chain, and applied to a lithiumniobate MZM under stable bias with a tunable C-band laser as optical source. OSNR is swept via ASE-based noise loading with optical attenuation and monitoring via optical spectrum analyzer and power measurements; the received waveform is acquired by a high-bandwidth photoreceiver and a real-time oscilloscope and stored together with OSNR tags and acquisition metadata. On this dataset, the thesis evaluates a compact one-dimensional convolutional neural network (1D-CNN) equalizer operating on fractionally spaced samples resampled to two samples per symbol. A fixed context window is processed by three convolutional blocks with increasing feature channels and regularization, followed by a small dense stage that outputs the two bits required to recover BPAM-4 symbols. A fractionally spaced linear FFE baseline (even/odd branches with different tap lengths) is used for comparison. The results show an OSNR advantage on the order of 5 dB for the CNN at the target BER relative to the linear baseline, indicating that learning-based equalization can exploit residual nonlinear and memory effects in a bandwidth-limited, pre-amplified IM-DD chain. 8. Discussion, limitations, and perspectives The thesis establishes that strict modularity and preservation of end-to-end coupling are compatible when the surrogate is built on disciplined interfaces and a unified time base. Device-aware modeling is most impactful when it targets coupling-critical mechanisms, including converter reconstruction and static nonlinearities, bandwidth-limited analog responses, amplifier dynamics, and sampling and quantization effects that jointly determine the effective discrete-time channel observed by DSP. By treating impairments as parameterized blocks and emitting structured metadata, the surrogate supports controlled sweeps, ablations, and dataset generation with traceable provenance. The current scope prioritizes IM-DD PAM and BPAM and a specific experimental configuration. Comprehensive multi-device calibration, formal uncertainty quantification, and broader format coverage remain future work. A near-term direction is full calibration and upgrade toward a digital shadow, meaning a measurement-aligned surrogate systematically updated as operating conditions and device settings vary. Further priorities include extending the impairment library (for example, richer converter and modulator models and mismatch diagnostics) and strengthening validation protocols that quantify domain shift and uncertainty when deploying models trained on synthetic or partially calibrated data. Mid-term perspectives include extending the framework to higher-fidelity laboratory alignment for multi-actuator transmitters, such as multiple electro-optic modulators and dual-polarization operation. In the long term, the goal is to upgrade toward a full digital twin of the IM-DD system by coupling device-based models with continual calibration, and to generalize the digital-surrogate methodology to coherent optical links and, where appropriate, to wireless communication setups using common transceiver building blocks. 9. Conclusion This thesis develops a device-aware, modular, end-to-end digital surrogate for short-reach IM-DD optical links and demonstrates its utility as both a physically meaningful simulation framework and a structured data generator for learning-assisted optimization. By exposing dominant impairments as controllable parameters and enforcing time, interface, and metadata discipline, the surrogate preserves critical end-to-end coupling while enabling reproducible studies, calibration-oriented refinement, and experimental demonstration of learning-based receiver DSP for bipolar PAM.