Physics-Informed Neural Network Modeling of Biodegradable Contaminant Transport through GCL/SL Composite Liners

Beneath the sprawling concrete and steel of modern landfills lies a quiet, critical defense system designed to protect the groundwater that communities rely on. This barrier is not a single wall, but a layered sandwich of materials, typically starting with a thick sheet of plastic and followed by a compacted layer of clay, and finally, a deep bed of soil. Its job is simple in concept but complex in reality: to catch toxic liquids, known as leachate, that seep from decomposing trash and stop them from poisoning the earth below. As this liquid moves downward, it carries dissolved chemicals that can be harmful. Over time, the soil and clay layers act as a filter, slowing the movement of these chemicals through a process called diffusion, where particles spread out from areas of high concentration to low concentration. In many cases, natural bacteria within the soil also break down organic pollutants, effectively eating the contaminants and reducing their danger. However, predicting exactly how fast these chemicals will travel, and whether the barrier will hold up over decades, is a difficult mathematical challenge. The layers have different thicknesses and properties, and the speed at which the liquid moves depends on how much pressure is pushing it from above. If the pressure is too high, the chemicals might rush through before the soil has a chance to clean them.

To solve this puzzle, a team of researchers has developed a new way to model these underground journeys using a type of artificial intelligence called a physics-informed neural network. Instead of relying solely on traditional computer simulations that break the soil into a grid of tiny boxes, this new approach teaches a computer program the fundamental laws of physics—how fluids move, how they spread, and how they decay—and then lets the program learn the specific behavior of the landfill liner. The researchers focused on a specific type of barrier made of a thin geosynthetic clay liner sitting on top of a much thicker soil liner. They tested two different versions of their AI model. The first version, a standard approach, asked the computer to guess the solution and then gently corrected it whenever it violated the laws of physics. The second version, a "hard-constrained" model, built the rules directly into the computer's thinking process, forcing it to obey the starting conditions and the boundaries of the system from the very first step.

The results of this study show that the hard-constrained model is significantly better at predicting the future of these contaminants. When the researchers simulated the movement of a common organic chemical, benzene, through the liner system under various conditions, the standard model struggled, particularly when the pressure of the liquid above the barrier was high. In those high-pressure scenarios, the standard model made errors that were nearly ten times larger than those made by the hard-constrained version. The improved model successfully captured the sharp, rapid changes in concentration that happen when contaminants are pushed quickly through the soil, a situation where the older method often stumbled. The researchers found that the best performance came when the computer used a specific mathematical function, known as a hyperbolic tangent, to smooth out its calculations, and when the network was built with a specific number of layers and connections. Too few layers left the model unable to grasp the complexity of the flow, while too many layers offered no extra benefit and made the calculations harder to finish.

Beyond simply predicting where the chemicals would go, the researchers also tested whether their model could work backward. In the real world, scientists often have very limited data; they might only know the concentration of a chemical at a few specific points in the soil at a few specific times. The goal was to see if the AI could look at these sparse clues and figure out how fast the soil was naturally breaking down the contaminant. The hard-constrained model proved capable of this task. When given data from a simulation where the chemical took ten years to lose half its strength, the model correctly identified that timeframe. It performed just as well when the breakdown took thirty or even one hundred years. The study also tested the model's resilience against noisy data, simulating the kind of measurement errors that happen in real-world monitoring. The model remained reliable when the data contained small amounts of noise, but its accuracy began to drop when the errors became very large, suggesting that while the tool is powerful, it still needs reasonably clean data to work its best.

This research offers a more precise and efficient tool for engineers designing landfill liners. By accurately predicting how long a barrier will last and how contaminants will behave under different pressures, communities can build safer landfills that protect groundwater for generations. The study confirms that embedding physical laws directly into the learning process of artificial intelligence creates a more robust and accurate predictor than methods that treat those laws as optional suggestions. While the current work focused on idealized, one-dimensional layers, the success of this approach suggests a path forward for tackling even more complex, real-world scenarios where soil properties vary and chemicals interact in complicated ways. The findings provide a strong foundation for using advanced computing to ensure that the invisible barriers beneath our feet remain effective long after the landfills above them are closed.