We propose a model-based machine-learning approach for polarizationmultiplexed systems by parameterizing the split-step method for the Manakov-PMD equation. This approach performs hardware-friendly DBP and distributed PMD compensation with performance close to the PMD-free case.