Machine Learning-Powered Predictive Analytics for Evaluating the Performance of Energized Foams

Deep beneath the Earth's surface, in rock formations so tight that oil and gas cannot flow freely, engineers must create artificial cracks to unlock these resources. This process, known as hydraulic fracturing, involves pumping fluid at immense pressure to fracture the rock and then pushing solid grains, called proppants, into those cracks to keep them open. The success of the entire operation hinges on the fluid's thickness, or viscosity. If the fluid is too thin, the proppants sink before reaching the end of the crack, leaving the reservoir unproductive. If it is too thick, it becomes difficult to pump. In the hottest and most pressurized environments, where temperatures can exceed 120 degrees Celsius and pressures reach over 3,000 pounds per square inch, predicting how this fluid will behave is notoriously difficult. To solve this, researchers have turned to "energized foams," which are mixtures of gas and liquid that use less water than traditional fluids while carrying the necessary solids. However, these foams are unstable; they can break down or change thickness rapidly under the extreme heat and pressure of a well, making it hard to know how they will perform until it is too late.

A team of researchers at the University of Adelaide has developed a new way to predict how these energized foams will behave, combining real-world laboratory data with computer models that learn from experience. Instead of relying solely on expensive and time-consuming physical tests, the team built a digital framework that can forecast the fluid's thickness as it changes over time. They gathered data from eleven distinct laboratory experiments involving carbon dioxide and nitrogen foams, capturing nearly 4,000 individual measurements. These tests covered a wide range of conditions, from pressures of 947 to 3,040 psi and temperatures from 25 to 122 degrees Celsius. The researchers focused on a specific type of foam stabilized by a combination of surfactants, which are chemicals that help mix oil and water, and silica nanoparticles, which are tiny solid particles that reinforce the foam's structure. By feeding this data into machine learning algorithms, they created two complementary tools: one that predicts the fluid's thickness based on the current conditions, and another that tracks how that thickness evolves as the experiment progresses.

The first tool, designed to understand the fluid at a single moment, successfully identified that temperature is the most critical factor. When the heat rises, the foam thins out, and the computer model learned to anticipate this drop with high accuracy. Among the various mathematical approaches tested, a type of neural network known as a multilayer perceptron proved to be the most reliable for this task. It could look at the current pressure, temperature, and chemical makeup of the mixture and estimate the fluid's thickness with a level of precision that rivals physical testing, but without the need for a new experiment every time. This model acts as a starting point, giving engineers a baseline expectation of how the fluid should behave before it enters the complex, changing environment of a well.

The second tool addresses the fact that foam does not stay the same; it degrades and changes as it is pumped and heated over time. To capture this, the researchers built a dynamic model that looks at the history of the fluid's behavior. Just as a person might feel tired after a long day of work, the foam's structure weakens after being subjected to stress. The computer model learned to recognize these patterns by looking at the fluid's thickness in the moments immediately preceding the current moment. The most effective algorithm for this job was a method called Lasso regression, which is particularly good at finding simple, stable patterns in data without getting confused by noise. This model successfully predicted the general trend of the fluid thinning out over time, smoothing out the jagged peaks and valleys that often occur in raw data to provide a clear picture of the fluid's lifecycle.

While the models performed well when tested on data similar to what they had seen during training, the researchers found that their accuracy dropped when faced with completely new experiments that they had never encountered before. This is a crucial finding, as it highlights the limits of the current approach. The models are excellent at interpolating, or filling in the gaps, between known conditions, but they cannot yet predict with perfect certainty how a completely different foam formulation or a drastically new set of conditions will behave. The study explicitly rules out the idea that these computer models can replace physical testing entirely. Instead, the researchers position their work as a powerful screening tool. It allows engineers to test hundreds of different scenarios on a computer to see which ones look promising, narrowing down the options before committing to the costly and hazardous process of physical laboratory testing.

The study also clarified what drives the fluid's behavior. In the static, moment-to-moment predictions, temperature was the dominant force, followed by the speed at which the fluid was being pumped. In the dynamic, time-based predictions, the most important factor was simply how the fluid had behaved in the recent past. This confirms that the history of the foam is just as important as its current state. The researchers noted that their data did not include continuous measurements of every possible variable, such as the exact amount of salt in the water or the precise concentration of nanoparticles at every second, which limits how far the models can be pushed.

Ultimately, this work provides a bridge between the messy reality of fluid dynamics and the clean precision of computer prediction. It offers a way to visualize how an energized foam will thin and change as it travels from the surface down into a hot, high-pressure reservoir. By combining a snapshot of the current state with a memory of the past, the system gives engineers a more complete picture of the fluid's performance. While it is not a crystal ball that can predict the future of any possible foam under any condition, it is a robust tool for making better decisions today. It helps prioritize which chemical mixtures deserve further study and which operating strategies are likely to succeed, saving time and resources while improving the safety and efficiency of extracting energy from the Earth's most challenging reservoirs.