Adaptive and scalable optical systems employing reconfigurable optical add-drop multiplexer (ROADM) units and flex-grid dense wavelengthdivision multiplexing (WDM) techniques have been enabling various applications and services that require high capacity and low latency in the underlying optical network. The erbium-doped fiber amplifier (EDFA) is a key hardware component that has been widely deployed in optical transmission systems, and can have a large impact on the end-to-end system performance such as the optical signal-to-noise ratio (OSNR) and quality of transmission (QoT) [1] , which depends on the power level of individual wavelength signals. Characterizing the gain spectrum profile of an EDFA is challenging since it is not only a complex function of many parameters such as the gain setting, channel loading configurations, and input power levels, but also a hardware-specific property of individual EDFA components. Since a reconfigurable end-to-end optical link through a mesh ROADM network may include multiple EDFAs of different types (e.g., preamp, booster, and in-line), the multiplicity of system configurations complicates the collection of corresponding end-to-end datasets. Through the use of predeployment lab data collection, component-level EDFA gain spectrum modeling for use with QoT estimation methods can provide potentially efficient and scalable power evolution prediction in such multi-span systems.
Recent work has focused on the gain spectrum modeling and optical channel power prediction of EDFAs using both analytical models [2] and machine learning (ML) approaches [3], [4] . It has been shown that ML-based EDFA models using neural networks trained on large measurement datasets can achieve accurate component-level modeling. Multi-span systems with multiple ED-FAs have also been considered, with a focus on power evolution and OSNR prediction [5], [6] . In particular, recent work [5] used end-to-end data collection and showed that using separate amplifier models can provide accurate end-to-end results. However, this work did not consider separately characterized EDFAs and used bench-top optical spectrum analyzers (OSAs) to measure the spectrum. Separate data collection using the built-in optical channel monitoring (OCM) capabilities of the ROADM units would allow for flexible in or out of system characterization and re-training without the need of extra bench equipment. However, built-in OCMs are less accurate and need to be studied for use in this application.
In this paper, we study optical signal spectrum prediction of a multi-span system consisting of ROADM nodes constructed using 95 channel, separately characterized Lumentum ROADM-20 whitebox units, deployed in the programmable COSMOS testbed [7] . For each EDFA in the multispan system, we collect a comprehensive set of power spectrum measurements under diverse channel loading configurations using the in-line OCMs that are built into the ROADM-20 whitebox units. Using the separately collected datasets, we develop component-level EDFA gain spectrum models using deep neural networks (DNNs), which predict the output power spectrum based on the channel loading configuration and input power spectrum. Transferring such individual models to a collective multi-span system, extensive experiments with diverse channel configurations show that the trained DNN-based EDFA model can accurately predict the power spectrum after 10 EDFAs with a mean absolute error (MAE) of 0.73 dB and 0.61 dB with two 5-span metroscale configurations, respectively.