[CCoE Notice] seminar at 2 pm, Tuesday Feb. 18, T2 Room 386: A Physics Model Embedded Hybrid Deep Neural Network for Drillstring Washout Detection

[Knudsen, Rachel W]

A Physics Model Embedded Hybrid Deep Neural Network for Drillstring Washout Detection

Tuesday, February 18, 2020

2 pm, Room 386, the Technology Building (T2)

Abstract: The industrial world is undergoing a transformation, enabled by new machine-learning techniques and increasingly large volumes of digital data.  Many of these data analytic methods work well when fed with a sufficient quantity of properly conditioned data, and the Oil & Gas industry has many challenges that can be addressed in this way.  However, there are several classes of problems where the lack of quality data in sufficient quantities precludes these approaches.

One way to tackle these data-poor problems is to build hybrid systems that combine models of physics with the data that are available. Ideally these systems should be able to take advantage of the many advances that machine-learning and data analytics have brought in recent years, while still leveraging the decades (and sometimes centuries) of investment in knowledge and effort incorporated in the models.

The solution presented here, a hybrid deep neural network (hybrid-DNN), embeds physics models into a machine learning framework that can be trained by the available data.  The proposed hybrid-DNN is composed of three components – ParameterNet for estimating model parameters, ResidueNet for predicting regression or classification results, and the physics model(s) appropriate for the problem at hand.  ParameterNet learns the system behavior based on the embedded physics model, which it controls through adjusting model parameters.  ResidueNet utilizes the outputs from the model and ParameterNet and trains a neural network to characterize the patterns and latent information in the derived parameters and data. Once trained the combined system can predict the desired results and indicators based on a real-time data input stream.

As a practical demonstration of the methodology, the physics model embedded hybrid-DNN has been applied to the problem of drillstring washout detection.  When fed with a real-time data stream it is able to extract best-fit model parameters and identify normal conditions and potential washout situations. As a general-purpose framework, the proposed method is applicable to many different domain problems based on physics.

Presenter’s Bio: Charles [Cheolkyun] Jeong is currently working for Schlumberger as an AI engineer at Katy Drilling Software Center (KDSC). Charles received his Ph.D. at Stanford University for integrated reservoir modeling with Bayesian statistics. His current role in well construction learning team is to develop integrated data analytics workflows and innovate industry AI-ML models.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://Bug.EGR.UH.EDU/pipermail/engi-dist/attachments/20200217/0136cb3e/attachment-0001.html