Atrial fibrillation (AFIB) and atrial flutter (AFL) require different clinical management, but their diagnostic distinction from short 12-lead electrocardiograms (ECGs) can pose a challenge. We propose an interpretable pipeline that combines a synthetic-data-trained 1D U-Net for atrial signal extraction with an autocorrelation-based distinction feature. A total of 50,000 synthetic 10-second ECGs were generated using cubic Hermite splines with diverse atrial and ventricular morphologies, that were further augmented by multiple noise profiles. The U-Net was trained on synthetic data to decompose each signal into atrial, ventricular and noise components, achieving Pearson correlations of 0.980 (atrial) and 0.998 (ventricular) on validation data. The trained model was then applied to 5,203 AFIB and 726 AFL 12-lead ECG recordings from the CinC challenge 2021 datasets in a leave-one-dataset-out cross-validation. The optimized distinction feature achieved a cross-dataset accuracy of 93.6%. Dataset-specific performance increased consistently over all datasets by up to +10.3% compared to the literature. Around 90% sensitivity, the method achieved 89.0% specificity compared to only 66.7% for a standard method. Cross-dataset and dataset-specific performance optimization showed similar findings, demonstrating generalizability and, thus, clinical usability in further applications.