A Physics-Informed ML Approach for Scalable Seismic-to-Well Tying Using Bulk Time Shifts

Seismic-to-well tying remains a critical yet time-intensive component of seismic interpretation workflows. While automated approaches have reduced manual effort, many rely on dynamic stretch-and-squeeze adjustments that maximize trace similarity at the expense of distorting the time–depth relationship. In large multi-well projects, maintaining both efficiency and geophysical consistency remains challenging. This paper presents a physics-informed machine learning workflow that jointly estimates the seismic wavelet and an optimal bulk time shift while explicitly avoiding dynamic warping of the time–depth relationship. The model is trained on a large synthetic dataset constructed from randomly generated reflectivity series convolved with analytical wavelets of varying phase and bandwidth, with known bulk shifts applied. Realistic seismic variability is introduced through controlled augmentation, including additive noise, smooth low-frequency distortion, multiples, phase perturbations, and nonlinear time warping. The workflow is demonstrated on multiple wells from the publicly available F3 Netherlands offshore field dataset. Results show improved temporal alignment at zero lag, preservation of velocity plausibility, and rapid scalability. The method illustrates how machine learning can be integrated into interpretation workflows while maintaining physical realism.