Share
CASE STUDY

GroundMetrics

Summary:
How has the storage reservoir of CO2 has changed over time
Tags:
“Expero was fundamental to the success on this project, effectively communicating their progress throughout and helping our scientists develop an understanding of deep learning techniques.”
Kris MacLennan
Principal Investigator at GroundMetrics

Challenges

Monitoring the status of CO2 reservoirs in Earth’s subsurface requires a team of highly trained specialists, running expensive computational resources for months at a time. The goal is to understand how the storage reservoir has changed over time as more CO2 has been added, and whether or not there are any leaks from the reservoir. The technological difficulties arise as a result of the “inversion” process of the electromagnetic signals used to visualize the reservoir. This inversion process can take months of computational runtime on expensive servers, requiring huge amounts of cap-ex and downtime to wait for results.


outcome

Alongside GroundMetrics, Expero co-authored a grant proposal aimed at solving the resource and time-intensive traditional approach to geophysical inversion for imaging CO2 reservoirs outlined in the Challenges section. When awarded the grant, GroundMetrics’ staff provided the geophysical tools and data, and Expero developed the data science components needed to solve the problem.


Our mixed team of geophysicists and data scientists solved the problem, designing, building, and tuning a system that output inversion results in less than a hundred milliseconds. This system eliminates the months-long inversion process requiring a huge computational cluster.


Business case

GroundMetrics customers needed a solution to the high cost, long cycle time process of monitoring CO2 reservoirs. A typical monitoring cycle costs a month or more of a geophysicist’s time and a month or more of a hundred core computational cluster. These resources can add up to $30,000 or more for a single cycle, required every time a customer wants updated visibility over one of their storage reservoirs. That direct cost, however, pales in comparison to the month-long cycle time. Customers have to wait a month for an updated view of each of their storage reservoirs.


This initiative reduced the computational time from a month to less than a second and doesn’t require a specialized electromagnetic geophysicist with deep knowledge of Maxwell’s equations to run.


approach

Interestingly, the solution to this problem had never been explored. Expero’s team undertook the exploration of this problem as a research project and ended up generating novel scientific outcomes as a result. A team of Expero data scientists architected, built, and tested several deep learning architectures to invert electromagnetic data, including several varieties of generative adversarial networks, variational autoencoders, and fully convolutional networks. After landing on a fully convolutional architecture with an encoder-decoder structure, Expero developed a specialized heuristic feature engineering approach that selects optimal feature transforms for input to the neural network.


We also invented a key component to high dimensional image work we’re calling Deep Dimensionality Exchange. Outlined in the cartoon image above, deep dimensionality exchange sits at the intersection of an encoder and a decoder. The DDE system transforms the number of elements in the input latent space into the number (and orientation) of elements in the output latent space, thus enabling domain transfer between input and output domains of any dimensions. We expect this to be critically important in all subfields of high dimensional imaging, like doppler ultrasound, satellite imagery, and hyperspectral analysis of defense targets.


user audience

  • CO2 sequestration engineers
  • CO2 storage systems geophysicists 
  • CO2 storage project managers


services

  • Data science R&D
  • Grant writing - SBIR
  • Deep learning model architecture
  • Distributed computing machine learning training
  • MLops

project details

  • 6 weeks of ML R&D
  • 3 person team
  • Agile experimentation process
  • Interactive collaboration with mixed client team
  • Co-authored grant acquisition