How Bluware Simplifies Deep Learning for Seismic Interpretation
There’s a global race underway to bring machine learning (ML) techniques to knowledge workers across every industry that extracts insights from data. Oil and gas is no exception. Many exploration and production (E&P) operators and service companies are exploring the technology to help illuminate patterns in seismic data where fuzzy pattern recognition is at the heart of many tasks. Fundamental innovations in ML during the past decade have led to an explosion of solutions to problems that have historically been too “fuzzy” to solve with traditional sequential code. ML could be a great fit for many of these challenging geophysical problems.
For example, many features like reefs, levees, and sand injectites aren’t amenable to traditional mapping technologies like auto-tracking or opacity carving, but their distinctive character in the seismic fabric makes them highly detectable with ML. In cases like this, it makes sense to bring deep learning into your workflow. The technology offers the potential to make you a faster, more productive interpreter and delivers a better quality solution than what you could do manually.
Until recently, deep learning-based workflows have been out of reach for the individual geoscientist. They required significant GPU compute resources and the aid of specialized data scientists and data engineers. Bluware, an AWS partner, and AWS have worked together to remove all these barriers of entry, democratizing access to deep learning for every geoscientist in E&P.
Machine Learning for the Geoscientist
Bluware InteractivAI™ is a software application designed to manage all the data science tasks so that geoscientists can apply deep learning to their seismic datasets. Training a neural network and making machine inferences is invoked by simply pressing a button.
InteractivAI’s deep learning engine depends on you to provide localized expertise about your data and basin. By learning from your interpretation, it suggests incrementally new interpretations as you make your way across a seismic dataset. As you reinforce the network’s inferences, the network will converge upon a comprehensive solution that mirrors your personal way of interpreting.
The combination of an expert’s judgement and the machine’s precision offer a potent new way to uncover subtle patterns in the subsurface and provide interpretation deliverables using less person-hours than ever before. We believe this collaborative approach between expert and technology will be the most effective path to harness machine insights when mapping the subsurface.
Enabling Technologies for Real-Time Machine Learning
To deliver a label-train-infer workflow work in real time, Bluware has combined a series of enabling technologies that have been encapsulated in a seismic storage format called Bluware Volume Data Store (VDS™).
VDS is a modernized approach to store and retrieve seismic data, designed to work equally well on your laptop at home or as an object-class store in a cloud-native environment. The traces are arranged in subsets to allow parallel reading, assuring to the traces when you make random pattern calls into the data.
VDS is also engineered for adaptive streaming. VDS can adapt the signal quality of the seismic to your use case. For visualization tasks, the traces can be streamed at a lower signal quality to achieve better latency performance. These same traces can then be streamed into a compute engine using the highest level of detail to assure precision.
Lastly, VDS uses compression to optimize transfer and storage performance. It’s not unreasonable to see SEG-Y datasets compressed to 25% of their original size in VDS format while retaining sufficient signal integrity for nearly every interpretation workflow. Even if someone in your organization has a principled objection to lossy compression on seismic data, a lossless compression option is now available while still maintaining support for adaptive streaming.
Nearly all of these technologies are freely available in an open-source format called OpenVDS available through the OSDU™ Data Platform. In October 2020, Bluware announced the offering of OpenVDS+, enabling our compression technology for everyone. OpenVDS+ is a free-to-use library that adds Bluware’s industry-leading wavelet compression technology to OpenVDS.
Regardless of which flavor of VDS is chosen, they can all be freely converted back to SEG-Y. By supporting open standards and sanctioning format conversion, Bluware wants to eliminate historical concerns of a vendor-locked format. OpenVDS and OpenVDS+ can yield benefits to the entire ecosystem of seismic data wranglers and not just a walled-off segment of our community.
Engineering a Machine Learning Data Pipeline for the Geoscientist
To deploy a fully self-contained deep learning environment focused on the seismic interpreter, Bluware built an entire data pipeline to coordinate the VDS seismic volumes, expert labels, and all of the neural network components.
Traditional deep learning workflows often start with a SEG-Y dataset and a set of interpretation labels. In the following figure, a data engineer extracts all the in-lines and cross-lines from both datasets and converts them to images. They are chopped up, randomized, and arranged into a record file. A data engineer then passes their work off to a data scientist who feeds the labels and seismic into a neural network to produce a trained network.
The process is always iterative, as the geoscientist must update the label set to cure some deficiency in the inference. This whole workflow can often consume weeks. InteractivAI has eliminated these individual steps so labels and traces can flow automatically into a neural network architecture directly from an interpretation environment, harnessing the power of VDS’s truly random access.
Just spending a few minutes labeling a dataset will be enough to return a useful inference that can guide and augment future labels as your project unfolds. This allows for the machine’s inference to start accelerating your interpretation immediately.
Running InteractivAI in the AWS Cloud
While InteractivAI enables a software-based solution to apply ML to your seismic, a powerful hardware platform is still required to power the code. InteractivAI was built using a web-based client/server architecture, using Amazon EC2 instances for GPU-intensive tasks like ML.
The geoscientist simply engages with the web application. The only end-user requirement for InteractvAI is an internet connection and a web browser. By separating the UI and the compute engine, even the most modestly equipped commodity laptops can harness the power of ML. This is a great real-world example of reducing capital IT budgets when moving workloads into the cloud.
Geoscientists communicate with the InteractivAI server on a G4 instance in Amazon EC2 in AWS. These instances have vast stores of GPU memory, a key requisite for enterprise-scale ML. As for data storage, VDS volumes are stored in Amazon S3 as an object, as opposed to a file on a much more expensive and slower traditional file system. Object class storage makes more technical and commerical sense for seismic datasets than traditional file-based systems.
There are a handful of options to get data into Amazon S3. . Your IT department can rent an appliance called an AWS Snowball, a petabyte-scale data transport solution, that can be used to transfer data to AWS. You can also establish an AWS Direct Connect, a secure and high-speed link between your data center and AWS.
All computations and data storage in InteractivAI happens on the server specified by your company’s IT department. Once the server portion of InteractivAI is running on AWS, it streams seismic data directly into your browser . Launching InteractivAI is as simple as typing in a web address into your browser.
ML-based techniques are finding their way into many data-rich industries. When it comes to seismic interpretation, Bluware’s InteractivAI™ was designed for geoscientists to deliver faster and more precise interpretations. It eliminates the laborious data preparation tasks that used to require data science staff support and provides a training pipeline that’s fast enough to enable interactivity. We are striving to build the tightest closed-loop ML system that blurs the line between your expertise and machine’s insights.
Are you interested in test driving InteractivAI on AWS at no charge to you? Receive one-on-one instruction and a hands-on test-drive on Bluware InteractivAI, powered by AWS.