04/11/2024

Pi Center uses neural networks to locate oil at sea.

The IMPA Center for Projects and Innovation ( Centro Pi ) is developing a project for Petrobras in the field of seismic inversion to increase the accuracy of oil exploration in deep waters. Still in its initial phase, the work innovates by using Physics-Informed Neural Networks (PINNs) to map the area being explored.

IMPA researcher João Pereira and project scientist Lucas Nissenbaum coordinate the team of eleven participants, including PhDs and postdoctoral fellows. Nissenbaum explains that the fundamental idea of the project is to help Petrobras analyze seismic data for oil exploration in a more assertive and rapid way. The ultimate goal is to avoid unnecessary drilling—which generates high financial and time costs—and optimize the work that already exists at Petrobras.

Ocean floor mapping for oil companies has been carried out using the Full Waveform Inversion technique since the 1960s. This is the most advanced and precise process for generating more specific models for this purpose. Despite this, a major challenge is that inverting fields using this technique demands a great deal of time and computational power, especially when dealing with complex or three-dimensional fields.

“One way to think about it is that the current working model starts with a guess about what is thought to exist at the bottom of the ocean. The better the guess, the faster the convergence of that model will be. The use of physics-informed neural networks could help in finding this first point, for example,” said Nissenbaum.

To achieve this, the Pi Center works in the field of seismic inversion. Using a jet cannon, a sound wave is generated, which travels and collides with different mediums. This process generates various effects on the wave, including its reflection at the surface. With the aid of sensors, the idea is to capture information about the mediums that affect this wave and thus estimate the composition of the field. With this analysis, it will be possible to perceive whether the medium is more or less dense in a certain region, for example, and produce an estimate of the amount of oil present.

“It’s a wave that travels across the earth, and its reflections generate new information. Fundamentally, we return to the thought of how this part of the earth was formed? During this formation, was it allowed to form petroleum? There are several aspects to be analyzed in this process,” explained Pereira.

The project coordinators highlight the Physically Informed Neural Network as a major focus of the developing model. The proposal is that these networks can obey classical physical models and properties, just as sound waves follow physical nature. The challenge is to teach, through machine learning, the neural network how to adapt to the physical context, to the real-world situation of the problem presented.

“The way we are presenting our model is by using neural networks. The function used in the project will be represented by a neural network, and that's what changes a lot in our research. There are several neural networks, and basically, we can see which ones work best and in what contexts. The way we train these neural networks to solve the equations changes. So, we can combine them to ultimately achieve something, to see what works best in the end,” Pereira pointed out.

Another important contribution so far has been the development of a pipeline for testing and evaluating results. This is a modular structure in which it is possible to test new ideas and models more easily. With the development of this tool, it has been possible to manipulate more complex and simpler equations, increase the size of the medium, and invert information and components.

“That’s what I think is the coolest thing about what we’ve done. Now, we want to carry out tests, replace one module with another. We have many research ideas, and we can do these tests quickly,” concluded Nissenbaum.