Euler Deconvolution (ED) is a cornerstone of semi-automated interpretation in potential-field geophysics, used to estimate the location and depth of magnetic and gravity sources. However, its practical utility is constrained by sensitivity to two user-defined parameters: the Structural Index (SI) and the moving window size. Traditionally selected through manual trial-and-error, this process is time-consuming, prone to subjective bias, and unreliable in noisy environments. This study introduces a hybrid framework that integrates ED with a 1D Convolutional Neural Network (CNN) predicting source depth and SI directly from normalized anomaly profiles. The network was trained on newly generated synthetic profiles of five geometries (dikes, cylinders, spheres, faults and sills) with 1–5% noise. A moving-window labelling of the optimal Euler window by the Coefficient of Variation was tested but found too noisy to be learned, so the window is fixed. On 1,000 independent held-out profiles, the final model, evaluated with Monte-Carlo dropout inference, gave a mean relative depth error of 5.0% (median 3.6%; 86% of profiles within 10%) and a mean absolute SI error of 0.14. Accuracy was maintained for noise up to 5% (mean errors of 4.1–6.7%) and degraded gradually at 10% and 15% noise (13.5% and 19.9%). The Euler step itself proved unstable under noise in our implementation, so the reported depths are the CNN depths. The framework was adapted to gravity by retraining on gravity forward models and applied to real profiles from the Ajichay region; these results are preliminary and sensitive to the regional-trend model. By removing manual parameter tuning, the method enhances the consistency and efficiency of geophysical interpretation.