Biosketch

Hongkyu Yoon is a principal member of technical staff at Sandia National Laboratories. He obtained a Ph.D degree in Environmental Engineering in Civil Engineering from the University of Illinois at Urbana-Champaign. After working as a research scientist at Illinois, he joined Sandia and his research focuses on subsurface carbon storage and energy recovery with emphasis on multiphase flow and reactive transport, chemo-mechanical coupling, and high-fidelity inverse modeling and uncertainty quantification. He is currently leading research projects on machine/deep learning development and applications for subsurface energy recovery, carbon storage, and remote sensing data analysis for greenhouse gas emissions in Sandia Earth Science program.

Introduction of the Lecture

Machine/deep learning (ML/DL) methods have recently emerged as promising methods for big data analysis and real-time forecasting of coupled subsurface processes. With traditional computational methods high dimensional forward and inverse problems for coupled subsurface processes have been challenged by the number of high fidelity forward model simulations and computational burdens with matrix calculations. Using ML-driven fast forward models and robust non-linear projection operators for dimension reduction, a ML-based framework can be developed as a robust and fast data assimilation solution for real-time forecasting. In this lecture a few examples of supervised and self-supervised ML applications for fast surrogate models will be first presented to highlight promising results of ML applications for coupled subsurface processes, and then generative models such as (variational) autoencoders and generative adversarial networks will be presented as non-linear projection operators to reduce the high dimensional parameter space to the low dimensional latent space. With well-based observation data the ML framework can be combined with a Bayesian inverse model or ensemble-based data assimilation to update the state model parameters for real-time forecasting. For demonstration purposes, poroelastic modeling of flow and pressure propagation in heterogeneous permeability fields and fractured media will be used to demonstrate the accuracy and speedup of the ML framework for real-time forecasting. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.

Biosketch

Dr. Yuxing Ben is a reservoir engineer at Occidental, where she develops hybrid physics and data-driven solutions in the subsurface engineering technology group. She was the principal developer of machine learning technology for Anadarko’s real-time drilling and hydraulic fracturing platforms. She won the best paper award from URTeC 2019 and was selected as a SPE distinguished lecturer for 2021. Prior to Anadarko, Dr. Ben served as the technical expert for Baker Hughes’ hydraulic fracturing software—MFrac. She has developed complex fracture model for Halliburton and was a postdoc at MIT. She has authored more than 30 papers and holds three US patents. She earned a BS in theoretical mechanics at Peking University, and a PhD in chemical engineering from the University of Notre Dame.

Introduction of the Lecture

This presentation will first introduce machine learning and its applications in oil and gas industry in the past few years, then share the experiences and learnings from three examples in real-time drilling and hydraulic fracturing.

• For real-time drilling, the operator developed a general machine learning model to classify rig states. Time series data was gathered from 40 wells with 30 million rows representing three US onshore basins. The model proved to have over 99% accuracy after being deployed on all the company’s unconventional drilling rigs. The model predicts real-time rig states every second with tolerant latency. The results are used to generate drilling KPIs in real time for drilling engineers in the office, aid in directional analysis, and optimize drilling operations.
• Continuous learning was used to predict wellhead pressure to avoid screenout and optimize completion costs in real time. More than 100 hydraulic fracturing stages were selected from several wells completed in the Delaware Basin. The wellhead pressure can be predicted with an acceptable accuracy by a neural network model. The ML model was tested in the Cloud, where real-time streaming data such as slurry rate and proppant concentration are gathered. The computation is fast enough that real-time wellhead pressure can be predicted.
• System identification was combined with model predictive control to allow the engineers to adjust the pumping schedule and optimize hydraulic fracturing costs.

The presentation will conclude with several takeaway points including future research and development directions for machine learning applications in oil and gas industry.

Biosketch

My research focus is on analyzing seismograms to understand how earthquakes work and to quantify the hazards they pose. My research interests include shallow earthquakes, intermediate-depth earthquakes, induced earthquakes, and slow earthquakes. We are working to improve earthquake monitoring in all settings by applying data mining and machine learning techniques to large continuous seismic waveform data. We also work on methods to anticipate the strength of shaking in earthquakes using the ambient seismic field. I currently serve as the Co-Director of both the Southern California Earthquake Center and the Stanford Center for Induced and Triggered Seismicity.

Introduction of the Lecture

Deep learning is having an impact across seismology, but its greatest impact has been in earthquake monitoring. The introduction of deep learning into earthquake monitoring workflows can reduce the detection threshold by one unit of magnitude or more, and because the number of earthquakes increases rapidly as magnitude decreases, detecting somewhat smaller earthquakes dramatically increases available information. The new generation of information-rich earthquake catalogs have yielded important insights into induced seismicity and hold the promise for yet more insights through the application of AI methods to the catalogs themselves.