Physics-guided machine learning for acoustic impedance inversion: a comparative study of MLFN and RBFN in data-constrained in an oilfield using rock physics-based pseudo-wells

Accurate prediction of subsurface acoustic impedance is a cornerstone of seismic reservoir characterization. Traditional model-based inversion techniques often fail under sparse well control and geologically heterogeneous conditions. This study introduces a hybrid workflow that integrates physics-based rock physics modeling with two machine learning approaches, namely multi-layer feedforward networks (MLFN) and radial basis function networks (RBFN), to enhance acoustic impedance inversion in a data-constrained clastic reservoir in the Persian Gulf. Rock physics modeling was applied to generate synthetic elastic logs, including compressional velocity, shear velocity, density, and pseudo-wells to enrich the training dataset and ensure geological plausibility. Seismic attributes representing amplitude, phase, and frequency characteristics were extracted and used as inputs for MLFN and RBFN architectures. A leave-one-well-out cross-validation strategy was employed to validate the models. Results show that both methods successfully reproduced acoustic impedance, with the MLFN achieving higher predictive accuracy, reflected in a cross-correlation of ۸۷%. At the same time, the RBFN offered faster training and robust performance in capturing localized nonlinearities. The comparative analysis highlights the respective strengths and weaknesses of both methods and emphasizes their complementary value for practical reservoir characterization. This workflow illustrates that physics-guided machine learning provides a reliable solution for generating impedance volumes while reducing reliance on dense well control, offering a promising approach for offshore and frontier exploration settings.

کلیدواژه ها:

seismic inversion ، acoustic impedance ، rock physics modeling ، machine learning ، MLFN ، RBFN