Atrial fibrillation (AF) is the most common arrhythmia, and detecting its fibrotic substrate is crucial for patient prognosis and optimal targeted therapies. Currently, non-invasive assessment relies heavily on late gad-olinium enhancement (LGE-MRI), a modality limited by reproducibility and clinical accessibility. Detailed 3D anatomical simulations with fi-brotic infiltration provide a robust computational alternative. By offering an exact ground-truth, they enable precise fibrosis prediction directly from clinically routine intracavitary electrograms (EGMs). Using 5 anatomically realistic anatomies with different degrees of fi-brosis infiltration, we simulated AF patterns and generated EGMs on a virtual endocardial shell (1 mm offset, N= 163,412 EGMs). Ground-truth fibrosis was defined for each EGM as the amount of fibrosis within a radius of 3mm. We have developed a preliminary deep learning method to predict presence/absence of local fibrosis (>5%) from temporally win-dowed (200 ms) unipolar EGMs: a 1D Convolutional Neural Network with 2 convolutional blocks, optimized to detect high-frequency frag-mentations and signal amplitude variance. Evaluated on the test cohort (32,002 EGMs, unseen virtual patient), the model achieved a Balanced Accuracy score of 0.73 and an AUROC of 0.81. This approach has the potential to predict high-resolution substrate mapping from clinical recordings, helping the personalization of digital twins for targeted ablation therapies.