Physics-Informed Explainable Machine Learning for Tight Sandstone Facies Prediction from Conventional Well Logs: A Blind-Well Comparative Study Krishna-Godavari Basin, INDIA

Deep beneath the earth's surface, in the vast sedimentary basins that hold the world's energy, geologists face a persistent challenge: finding oil and gas trapped in rock that refuses to let it flow. For decades, exploration focused on porous, sponge-like reservoirs where fluids move easily. Today, as those easy targets dwindle, the industry is turning its attention to "tight" reservoirs. These are rock formations where the pores are so small and the connections between them so narrow that the oil or gas is locked in place, requiring massive engineering efforts to extract. Identifying these tight zones is difficult because the tools used to look underground—long strings of sensors lowered into boreholes—often give confusing signals. A rock that looks like a good reservoir on a sensor reading might actually be a dense, impermeable block, and vice versa. This ambiguity makes it hard to know where to drill or where to spend millions of dollars to fracture the rock and release the energy.

In the Krishna-Godavari Basin on India's eastern coast, a researcher tackled this problem by teaching computers to see what human experts struggle to distinguish. They focused on a specific type of rock called tight sandstone, which sits deep underground and is mixed with shale, siltstone, and clean sand. The goal was to create a system that could look at standard sensor data from a well and accurately predict which type of rock it was passing through. To do this, they did not just feed the computer the raw numbers from the sensors. Instead, they first used the laws of physics to transform those raw numbers into new, more meaningful descriptions of the rock's behavior, such as how stiff it is or how it resists compression. They then trained several different types of learning algorithms to recognize the patterns of these rock types and tested the system on wells it had never seen before to ensure it wasn't just memorizing the answers.

The researcher worked with data from five wells in the basin, three of which had been carefully labeled by geologists who knew exactly which rock type was present at every depth. The other two wells were kept hidden, serving as a blind test to see if the computer could apply what it learned to a new location. The researcher compared two approaches. The first approach used only the standard measurements, such as the natural radioactivity of the rock, its electrical resistance, and how fast sound waves travel through it. The second approach took those same measurements and added a layer of calculated properties derived from physics, such as the rock's ability to store energy or its resistance to breaking. They tested four different machine learning models, including methods that build many decision trees and others that learn by correcting their own mistakes, to see which combination worked best.

The results showed a clear difference between the two approaches. When the models relied only on the standard sensor readings, they could identify the rock types with moderate success, achieving an average accuracy of about 66 to 68 percent. They were often confused by rocks that looked similar on the sensors but were actually different. However, when the models were given the additional physics-based descriptions, their performance jumped significantly. The average accuracy rose to between 81 and 84 percent. More importantly, the ability to correctly identify the specific "tight sandstone" zones, which are the most critical and difficult to find, improved dramatically. In the best cases, the models correctly identified these tight zones nearly every time, with a success rate exceeding 97 percent. This improvement was not a fluke; statistical testing confirmed that the gain was real and consistent, with the improvement in performance being so large that it was highly unlikely to have happened by chance.

To understand why the physics-based approach worked so well, the researcher used a technique that acts like a spotlight, revealing which specific clues the computer was using to make its decisions. They found that the standard sensor readings alone were often ambiguous, but the calculated properties provided a clearer picture of the rock's internal structure. For instance, the stiffness of the rock and how it responded to pressure helped distinguish a tight, dense sandstone from a clean, porous one, even when their basic sensor readings were nearly identical. The study also developed a way to measure how confident the computer was in each prediction. When the different models disagreed with each other, or when their confidence was low, the system flagged those sections for human review. This creates a safety net, ensuring that geologists focus their attention on the most uncertain parts of the well rather than trusting a computer blindly.

This work demonstrates that combining traditional sensor data with physics-based calculations creates a much more reliable map of the underground world. It suggests that the key to unlocking these difficult reservoirs lies not just in better sensors, but in better ways of interpreting the data those sensors provide. By teaching machines to understand the physical properties of the rock, rather than just the raw numbers, the industry can make more accurate predictions about where the oil and gas are trapped. The study concludes that this method is a powerful tool for characterizing these complex reservoirs, provided that the models are tested rigorously on new wells to ensure they are truly learning the geology and not just memorizing the data. The findings offer a clear path forward for exploring the deep, tight formations that will likely fuel energy needs in the coming decades.