AI Breakthrough Offers Fast and Accurate Predictions for Viscous Fingering in Complex Fluid Flows
Researchers have taken a major step toward solving a problem that has puzzled scientists and engineers for nearly a century: the unpredictable behavior of viscous fingering, a phenomenon that occurs when a less viscous fluid pushes a more viscous fluid through a porous material. This instability creates branching, finger-like patterns that are notoriously difficult to model, especially when the viscosity contrast between the two fluids is high. A new deep-learning framework developed at the USC Viterbi School of Engineering now provides a faster, more accurate way to understand and predict this behavior.
The team, led by associate professor Birendra Jha, collaborated with graduate students Ramdhan Wibawa and Mohammed Alasker to produce a system that dramatically outperforms traditional simulation methods known as Direct Numerical Simulations (DNS). DNS has long been the industry standard for modeling fluid mixing in porous media, but it comes with significant limitations. These models often require enormous computational resources and can take weeks or months to complete a single high-resolution simulation. In some cases, older simulations took around two months just to model one scenario — far too slow for real industrial needs.
The new AI-driven method changes that. Once trained, the deep-learning framework can produce results almost instantly — within the time it takes to run a simple query. The training process itself only required around an hour using two GPUs, a small fraction of what high-fidelity physics-based simulations demand.
At the core of this new system is a clever combination of two major techniques. The first is spatial embedding, executed through an autoencoder neural network. Autoencoders are powerful tools for identifying and compressing complex patterns, and in this case, the network learned to capture the detailed shapes and structures of fluid fingers as they form and evolve. The autoencoder reduces these detailed patterns into compact, meaningful digital signatures that still retain all the essential information. This method effectively recognizes the multi-scale patterns responsible for the complicated shapes of viscous fingering.
The second key component is the use of Koopman-based temporal dynamics, a mathematical strategy designed to simplify the prediction of how fluid patterns change over time. Instead of attempting to directly model highly nonlinear fluid motion — which is extremely difficult — the Koopman operator transforms the system into a different space where the dynamics behave more linearly. This allows the model to accurately forecast future states of the fluid, including events like the splitting, merging, and coarsening of fluid fingers, all of which are hallmark behaviors of viscous fingering.
This combination of methods resulted in a model capable of predicting multiple aspects of the phenomenon with very high fidelity. The researchers noted that, aside from matching the accuracy of DNS, the AI was actually able to eliminate certain errors that traditional simulations tend to produce. For example, in DNS, there may be small patches showing the presence of the more viscous fluid in areas where it should not exist — an artifact of numerical inaccuracies. The AI model removed such inaccuracies entirely, providing cleaner, more physically consistent predictions.
The breakthrough carries enormous implications for a variety of scientific and industrial fields. Viscous fingering is not just an academic curiosity — it affects major applications such as enhanced oil recovery, CO2 sequestration, and groundwater remediation. In each of these scenarios, understanding how fluids displace each other in porous environments is essential for optimizing operations and ensuring safe and efficient outcomes. Faster, more reliable modeling could help companies and researchers make quicker decisions, reduce costs, and improve environmental impact.
Beyond subsurface applications, the researchers emphasize that viscous fingering also appears in microfluidic devices, which are commonly used in biomedical and pharmaceutical work. Even though microfluidic channels are vastly different from underground rock, the fundamental physics of viscous fingering still apply. When fluids of different viscosities interact in these tiny systems, similar instabilities can occur, affecting experiments, drug delivery, and diagnostic processes. Being able to quickly and accurately simulate fluid behavior in such devices could be a valuable tool for biotechnology innovation.
The research team plans to expand the model’s capability by incorporating more training data that reflects a broader range of geological conditions and porous-media structures. Real subsurface environments are highly heterogeneous, containing diverse rock types, pore structures, and flow pathways. By feeding more varied data into the AI model, the researchers hope to build a system that generalizes well across many real-world conditions, ultimately making the tool even more practical for field-level predictions.
This development also fits into a broader trend in engineering: the increasing use of AI to accelerate computationally heavy simulations. Traditional fluid dynamics simulations, especially when they deal with nonlinear instabilities, multiphase flow, or fine-scale resolution, can become extremely demanding. AI-based surrogate models like this one offer a pathway toward near-instant simulations without sacrificing the accuracy needed for scientific or industrial confidence. As more researchers adopt machine-learning-based approaches, we’re likely to see many fields transition toward hybrid modeling, where physics-based equations and data-driven methods work together.
Understanding Viscous Fingering
Since this phenomenon may be unfamiliar to some readers, it’s useful to know a bit more about how it works. Viscous fingering typically appears in systems where two fluids interact inside a porous medium, such as a layer of packed sand, soil, or rock. When a low-viscosity fluid pushes a thicker one, the interface becomes unstable — the thinner fluid tends to penetrate in narrow pathways, forming finger-like structures. These fingers grow, branch, and merge in ways that depend on viscosity contrast, flow rate, diffusion, and the geometry of the porous material.
This instability is part of a broader category of fluid-mechanics behaviors known as interfacial instabilities, which also includes phenomena like the Saffman–Taylor instability in Hele-Shaw cells. Such patterns are influenced by nonlinear processes, making them historically difficult to capture accurately.
Understanding and controlling viscous fingering is important because it often causes inefficient mixing or uneven displacement. For example, in enhanced oil recovery, if water injected into a reservoir forms fingers instead of a smooth front, it may bypass large amounts of recoverable oil. In groundwater cleanup, fingers can carry contaminants into areas that would otherwise remain unaffected.
The Future of AI-Driven Fluid Modeling
With this new study, it’s clear the research community is moving toward merging machine learning with classic physics-based modeling. The ability to convert a computationally intensive problem into one that runs almost instantly opens the door to real-time monitoring, rapid scenario testing, and smarter decision-making.
As more datasets become available, particularly from real-world experiments, AI models may soon be able to predict fluid behavior across three-dimensional geological structures with unprecedented accuracy. For industries that depend heavily on subsurface fluid flow, this could lead to significant advancements in both efficiency and sustainability.
Research Paper:
Deep learning models of viscous fingering based on Koopman dynamics of dense embeddings
https://doi.org/10.1103/knp4-cd89
