Bluware Corp.

U.S. Headquarters

16290 Katy Freeway

Suite 400

Houston, TX 77094

Europe Regional Office



+1 713.335.1500

Email Us

Karl Johans gate 27

0159 Oslo, Norway

  • Twitter - White Circle
  • LinkedIn - White Circle
  • Vimeo - White Circle


Accelerate Machine Learning Using Interactive Deep Learning



Bluware DEEP LEARNING is interactive and driven by Geoscientists. This case study demonstrated what can be achieved in two hours compared to three months using traditional techniques.

No, Bluware can make a big difference way beyond seismic data.

It so happens that feeding data into computation and machine learning has become an increasingly large bottleneck where GPUs and TensorFlow processing units are starved for data, because the speed of IO and CPUs grow slower than the speed of GPUs.

To mitigate this issue, Nvidia  has released DALI, a library containing both highly-optimized building blocks and an execution engine for data pre-processing in deep learning applications, proving that they can train deep learning benchmarks faster than anyone else. The outcome is that DALI is no more than a small subset of the current Bluware engine.

For example, DALI uses hardware JPEG file decompression to bypass the slower CPU by feeding the GPU with images, whereas Bluware uses its proprietary compression technology and data augmentation engine to perform on-the-fly data format conversion. Bluware applies other augmentation filters that would otherwise require lengthy pre-processing and expensive data duplication.


Using the Bluware engine to ingest ImageNet data (1 million images of different categories) into the Bluware format, resulted in a 50 percent smaller dataset than DALI [~4 decibels better Peak Signal to Noise Ratio (PSNR)]. It was also two-times faster to load all the images in Python (reaching more than 3GB per GPU), leading to two-times better training performance and two-times better initial data compression.

Bluware’s data format is optimized to be streamed directly from an Amazon Simple Storage Service (S3) object store in a server-less way from any machine learning framework, compute, or visualization application.

Since these capabilities and results were achieved in such a different domain from oil and gas seismic imaging, it confirms that the Bluware engine can easily be evolved to play in other areas where signal data is used, such as IoT, medical imaging, or autonomous vehicles.

An oil and gas exploration company can setup a data science organization working on several artificial intelligence projects including the setup of a machine learning infrastructure. Data scientists can develop machine learning networks training on masses of data. The result is a machine learning service that can be utilized by asset teams. In this machine learning service, seismic data must be submitted for the service to convert 3D seismic into images for TensorFlow records. This portion of the workflow can be completely augmented with the Bluware Machine Learning solution that enables interactive deep learning and inferencing.

Companies have spent significant resources in the last several years to develop a supervised machine learning (ML) capability with varying degrees of success. Unfortunately, geoscientists are already battling uncertainty in their daily interpretation tasks. Anything that diverts them from the real-world issues is not an option for deployment into real projects and asset teams.

Several machine learning implementations have reached accuracy levels that approach 60 to 80 percent, but for a domain that is already dealing with data that is perhaps a 60 to 80 percent representation of the subsurface reality, the result is 36 to 64 percent accuracy, compared to the work completed by senior-level geoscientists.

The implications of “missing the mark” are not acceptable. Bluware delivers the ability to integrate machine learning directly into seismic interpretation, enabling geoscientists to drive and train the machine learning network while interpreting in real-time.


Bluware’s unique high-speed random data access gives an interactive experience, so that a geoscientist has full control of training the machine. In a few minutes of working on 8 to 10 in-lines or x-lines, the complete seismic data set is interpreted with greater than 99 percent accuracy.

Bluware has only applied this to post-stack seismic data, but it can also be applicable to pre-stack seismic data. Applying this to pre-stack seismic data will further accelerate the entire workflow, removing an expensive step in the seismic interpretation process. Bluware currently provides two machine learning networks including fault and salt extraction.



Tim Roden, Shell GeoSigns Software Manager

Our machine-learning can process seismic data to find geologic faults faster. Depending on the geology, some fault lines can help oil and gas migrate to the surface while others can disrupt drilling operations. Machine learning allows him to do that work in two hours when it used to take geologists two months through pore through the data. That’s transformational.


Interested in learning more about Bluware DEEP LEARNING? 

Schedule a meeting with us.