## SAGE — Subsurface foundational model with AI-driven Geostatistical Extraction

speaker: [Felix J. Herrmann](https://slim.gatech.edu/people/felix-j-herrmann)  
affiliation: Seismic Laboratory for Imaging and Modeling ([SLIM](https://slim.gatech.edu)), Georgia Institute of Technology

**Abstract.** *In this talk, we introduce a novel diffusion-based generative modeling approach for synthesizing high-fidelity subsurface velocity models. Once trained, this foundational model serves as a prior or a generative tool, enabling the sampling of velocity models that closely match the statistical distribution of geologically relevant Earth models. To achieve this, we leverage geostatistical information derived from curated datasets provided by the UK National Data Repository. Our training set comprises check-shot-corrected migrated images intersected by well logs. Numerical experiments on multiple synthetic datasets, including the Compass model and Synthoseis, demonstrate the model’s ability to generate realistic subsurface velocity models. This capability offers a data-driven prior for full-waveform inversion and facilitates the generation of training samples for amortized neural networks to improve full-waveform inference.*

This work is conducted in collaboration with [Huseyin Tuna Erdinc](https://slim.gatech.edu/people/huseyin-tuna-erdinc) and [Thales Souza](https://slim.gatech.edu/people/thales-souza).



<!-- ## SAGE -- Subsurface foundational model with AI-driven Geostatistical Extraction

speaker: [Felix J. Herrmann](https://slim.gatech.edu/people/felix-j-herrmann)  
affiliation: Seismic Laboratory for Imaging and Modeling ([SLIM](https://slim.gatech.edu)), Georgia Institute of Technology

**Abstract.** *During this talk, we present a novel approach to synthesize complex high-fidelity subsurface velocity models using diffusion-based generative models. After training, this foundational model can be used as a prior or as a generative tool to sample velocity models from a distribution that is close to the statistical distribution of relevant Earth models. To this end, our method is designed to capture geostatistical information from curated data obtained from the UK National Data Repository. The training data consists of check-shot corrected migrated images intersected by well-logs. Numerical experiments conducted with multiple synthetic datasets (Compass model and Synthoseis) suggest that our approach facilitates realistic subsurface velocity synthesis, providing a "data-driven" prior for full-waveform and an ability to draw samples that can be used to train amortized neural networks to conduct full-waveform inference.*

This is joint work with [Huseyin Tuna Erdinc](https://slim.gatech.edu/people/huseyin-tuna-erdinc), [Rafael Orozco](https://slim.gatech.edu/people/rafael-orozco), and [Thales Souza](https://slim.gatech.edu/people/thales-souza) -->

------------------------------------------------------------------------

<!-- ## Towards a pressure sensitivity-aware Digital Shadow for Geological Carbon Storage

speaker: [Felix J. Herrmann](https://slim.gatech.edu/people/felix-j-herrmann)  
affiliation: Seismic Laboratory for Imaging and Modeling ([SLIM](https://slim.gatech.edu)), Georgia Institute of Technology

**Abstract.** *We present an uncertainty-aware Digital Shadow for Geologic Carbon Storage (GCS) capable of handling pressure-induced changes in the compressional wavespeed. In order to incorporate pressure-dependence into our Digital Shadow for GCS, we make use of sensitivity-aware amortized Bayesian inference (SA-ABI), a recently developed statistical technique to integrate sensitivity analyses into simulation-based inference with neural networks in a computationally efficient manner. To this end, we exploit weight sharing to encode structural similarities between simulations based on different likelihoods---i.e., rock-physics models based on different effective stress coefficients, $n$, which appear in the equation for the effective stress, $\sigma^\prime$. This coefficient determines to leading order how the compressional wavespeed responds to changes in the pressure. With this approach, we leverage fast amortized inference by neural networks used to assess sensitivity to changes in the effective stress coefficient. Contrary to most Bayesian inference methods, our approach avoids the costly bottleneck of retraining the neural networks for each choice of $n$. This approach will also allow us to ask the network to produce samples for the inferred state---i.e., samples from the posterior distribution conditioned on observed seismic data, for different effective stress coefficients. With this approach assessments can be made on the robustness of the Digital Shadow to changes in leading-order parameters for the rock physics.*

This is joint work with 
Ipsita Bhar, [Abhinav Prakash Gahlot](https://slim.gatech.edu/people/abhinav-p-gahlot), and [Huseyin Tuna Erdinc](https://slim.gatech.edu/people/huseyin-tuna-erdinc) -->

## Towards a Pressure-Sensitivity-Aware Digital Shadow for Geological Carbon Storage

speaker: [Felix J. Herrmann](https://slim.gatech.edu/people/felix-j-herrmann)  
affiliation: Seismic Laboratory for Imaging and Modeling ([SLIM](https://slim.gatech.edu)), Georgia Institute of Technology

**Abstract.** *We present an uncertainty-aware Digital Shadow for Geological Carbon Storage (GCS) that accounts for pressure-induced changes in the compressional wavespeed. To incorporate this pressure sensitivity, we apply sensitivity-aware amortized Bayesian inference (SA-ABI)---a recently developed statistical technique that efficiently integrates sensitivity analyses into simulation-based inference using neural networks. Our method leverages weight sharing to capture structural similarities between simulations derived from different rock physics models, each characterized by distinct effective stress coefficients, $n$, which appear in the equation for effective stress, $\sigma^\prime$. This coefficient governs, to leading order, the compressional wavespeed's response to pressure changes. By utilizing fast amortized inference, our neural networks can evaluate the impact of variations in the effective stress coefficient. Unlike most Bayesian inference methods, our approach eliminates the computational bottleneck of retraining the neural network for each value of $n$. Additionally, the network can generate samples of the inferred state---posterior samples conditioned on observed seismic data---for different effective stress coefficients, enabling robustness assessments of the Digital Shadow with respect to these key rock physics parameters.*

This work is conducted in collaboration with Ipsita Bhar, [Abhinav Prakash Gahlot](https://slim.gatech.edu/people/abhinav-p-gahlot), and [Huseyin Tuna Erdinc](https://slim.gatech.edu/people/huseyin-tuna-erdinc).