Deep learning for predicting behaviors is becoming indispensable in the oil industry. Understanding the fundamentals and having a hands-on experience with hacking Python code in order to predict reservoir flow was the experience provided by the first AAPG-Halliburton Hackathon, which took place July 19 in Houston. Led by the AAPG’s Deep Learning Technical Interest Group (TIG) and Halliburton, the event attracted more than 120 registrants who were primarily geoscientists, engineers, and data scientists who worked together in teams that competed against each other and were judged by a panel of experts. The hackathon was powered by the OpenEarth^® Community and supported by a team of Halliburton experts in Python software, code, ideation, and strategy – as well as by domain experts and judges from the Deep Learning TIG.

Welcome to an interview with Andrew Munoz and Patrick Ng, co-chairs of the AAPG Deep Learning Technical Interest Group.

1.What was the full name of the Hackathon, and who was it intended for?

AM: Despite its seemingly daunting name, the “Domain Meets Deep Neural Networks: Hybrid Physics-based Hackathon for Geoscientists and Engineers” was intended for all skill ranges. Halliburton did all of the heavy lifting when it came to the creation of this deep learning neural network code, and we were the fortunate ones who got to play with it!

PN: It was called the “Domain Meets Deep Neural Networks: Hybrid Physics-Based Hackathon for Geoscientists and Engineers” Hackathon, and it was intended for domain geoscientists, engineers and decision makers, as well as data scientists who work as part of a team. Hence, it could be described in short as the Halliburton-AAPG Collaborative Hackathon.

2.What does the “physics-based” mean?

AM: Physics-based implies that the data model, which is used to predict some pattern using deep-learning neural networks (DNNs), has constraints based on physical models. These constraints are driven by analytical equations that have been proven in experimentation. Essentially, we are trying to “ground-truth” the black box that is neural networks.

PN: This physics-informed deep neural network uniquely combines the physics from data and machine learning.

3.What is a “deep neural network” as opposed to a regular neural network?

AM: A neural network is composed of a “hidden” layer (giving it its mysterious black box qualities) that takes in multiple variables and maps them non-linearly to predict some outcome. A deep neural network is simply a neural network with multiple cascading layers. It is a layer-upon-layer system, if you will, allowing for more complex feature extraction. For instance, a neural network might be trained to tell the difference between a man and a woman in a picture; however, a deep neural network can extract more nuanced information, like a brand of clothing that those same people are wearing. It simply extracts more features than a single-layer neural network when training on a data set.

PN: Deep means more layers, say more than three.

4.How did the Hackathon work? What were the problem sets? Where did the data come from? Where did the programming take place?

AM: The Hackathon started early Thursday morning, with a group of engineers, geologists, geophysicsts, data scientists, and computer scientists all gathered together to learn about the machine learning, Python, and Halliburton-provided recursive neural network template. The problem sets provided included a pollution dataset and a multi-well production dataset from a Norwegian platform.

PN: The Hackathon format had a pre-event webinar at the start to review Python essentials. The problem sets were simulated, and Halliburton provided actual reservoir flow (or production) data. Programming and hands-on coding took place at the Hackathon. Within each team, it was all hands on deck.

5.How did the teams form?

AM: Teams were initially assigned at the beginning, but we all shuffled around and ended up naturally being well distributed by the organizers of the event. My team comprised myself (geophysics), a computer scientist, a reservoir engineer, and a data scientist.

6.What did the judges do? What were their criteria?

AM: At the end of the hackathon, the teams presented their work via a formal presentation to the judges and the other participants. The judges primarily looked for creative approaches to using the machine learning code provided, along with any hacks that the teams came up with to improve the code or to see a different result.

PN: Judging criteria was provided by the organizers as follows (with points):

• Technical complexity (20)
• Prediction accuracy (20)
• DNN architecture (20)
• Geological legitimacy (or reasonableness) (10)
• Originality (10)
• Usefulness (10)
• Presentation of the results (5)
• Computational time (5)

7.What were some of the main lessons learned?

AM: First, real data is very difficult to clean. Many teams struggled with checking the validity of the data and running diagnostics. Second, running complex machine learning algorithms was easy with Python! Many Python libraries contain sophisticated algorithms that are readily accessible to the most basic Python user. This makes playing in Python much easier. Finally, everyone who participated looked at the problem sets differently, so coming together with great minds from other companies and disciplines helped to stimulate new and interesting ideas.

PN: 1) It was challenging to find the right mix of background within a team. 2) Pre-event bootcamp could extended to support individualized questions. 3) Time management –More coding time could benefit participants for future hackathons.

8.Do you have any additional thoughts that you would like to add?

AM:I’m looking forward to the next AAPG hackathon! We certainly have big plans for the next one.

PN: The OpenEarth.community provisions rapid experimentation through it’s DataScience Environment to all participants allowing creativity to flourish throughout the hackathon.