| Abstract Submission No. | ABS-4.1-0044 |
| Title of Abstract | Rock Physics Guided Variational Bayesian Neural Network for Elastic Log Prediction |
| Authors | Vivek Thomas*, Shib S Ganguli*, Prakash Kumar |
| Organisation | CSIR-National Geophysical Research Institute |
| Address | National Geophysical Research Institute Hyderabad, Telengana, India Pincode: 500007 E-mail: vivekthomasgeo@gmail.com |
| Country | India |
| Presentation | Oral |
| Abstract | Variational Bayesian Neural Networks (VBNNs) have emerged as a powerful class of models that provide a principled way of integrating prior domain knowledge with data-driven learning, thereby enabling robust predictions with predictive uncertainty. This helps to address the non-uniqueness of geophysical problems and the limited availability of labelled data. Reliable estimation of elastic properties such as P-wave (Vp) and S-wave (Vs) velocities is fundamental for seismic reservoir characterization. However, obtaining high-quality elastic logs is often constrained by acquisition cost, measurement noise, and sparse sampling. To address these challenges, we present a VBNN augmented with a rock physics workflow (RPW-VBNN). The idea is to embed meaningful constraints into deep learning by supplementing limited training data with synthetic elastic logs generated using effective medium models. These synthetic logs capture plausible velocity trends under varying lithological and petrophysical conditions, thereby guiding the learning process toward geologically consistent solutions. The resulting workflow is designed to provide accurate predictions, quantify uncertainty, and respect the underlying physics of the subsurface. Our approach is trained on petrophysical well logs from the Theia 1 well and validated on an independent blind well from China. The results clearly demonstrate the advantages of incorporating rock physics priors. Compared to a standard VBNN, the RPW-VBNN reduces mean absolute error by 73.1% for Vp and 61.9% for Vs. Furthermore, the CC significantly improved from 0.905 to 0.948 for Vp and from 0.896 to 0.963 for Vs. These highlights both the proposed method's predictive accuracy and enhanced reliability. The key contribution of this study is the demonstration that coupling rock physics knowledge with Bayesian deep learning leads to more trustworthy elastic property estimation, particularly in data-sparse or noisy environments. The RPW-VBNN framework provides a pathway toward more interpretable and geologically sound predictions by effectively bridging data-driven inference with physical modeling. |
| Keywords | Bayesian Neural Network, Deep Learning, Rock Physics, Elastic Logs, Reservoir Characterization |