📄 other

A Physics-Informed Computational Framework for Forward Prediction and Inverse Closure Learning of Pore-Scale Biofilm Growth

This study presents a physics-informed computational framework that integrates a modified density-dependent biofilm diffusion model with microfluidic observations to achieve superior forward prediction and inverse closure learning of pore-scale biofilm growth in porous media.

Original authors: Qiang Ye, Guangfeng Zhu, Qiang Zheng, Yu Wu, Lingzao Zeng

Published 2026-07-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Qiang Ye, Guangfeng Zhu, Qiang Zheng, Yu Wu, Lingzao Zeng

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a microscopic world hidden inside a sponge, a rock, or even a water filter. In this tiny universe, invisible communities of bacteria build sticky cities called biofilms. These aren't just static blobs; they are living, breathing structures that grow, eat, and move. To understand them, scientists use microfluidics, which is like building a transparent, miniature maze out of glass. This allows them to watch the bacteria in real-time as they drink nutrients and expand, much like watching a time-lapse video of a city being built. However, predicting exactly how these cities grow is incredibly hard. It's a "multiphysics" puzzle, meaning you have to solve for how water flows, how food (nutrients) spreads, and how the bacteria multiply all at the same time. If you get the math wrong, your prediction might look like a smooth, perfect blob, while the real thing is a messy, lumpy, and unpredictable structure. Getting this right matters because biofilms can clog water pipes or help clean up oil spills, so knowing how they behave is crucial for engineering and the environment.

Now, meet the team of researchers who decided to teach a computer how to solve this messy puzzle better than ever before. They didn't just let the computer guess based on pictures, nor did they just rely on old-school math equations that sometimes miss the mark. Instead, they built a "Physics-Informed Neural Network" (PINN). Think of this as a student who is studying for a test. A purely data-driven student just memorizes the answers from the textbook (the experimental photos). A traditional math student tries to solve the problem using only the rules of the universe (the physics equations). But this new AI student is a hybrid: it memorizes the photos and is forced to follow the rules of the universe at the same time.

The researchers started by watching bacteria (Bacillus subtilis) grow in their glass maze for 69 minutes. They noticed something surprising: the bacteria didn't just grow faster and faster as they got bigger. Instead, their growth followed an "S-shaped" curve. At first, they grew slowly, then they exploded in size, but eventually, they hit a wall. It was like a party where the room gets so crowded that people stop dancing and just stand still; the bacteria ran out of space and nutrients, and their own "self-inhibition" (a fancy way of saying "we're too crowded to grow anymore") kicked in.

To capture this, the team created a new computer model. They used a "physics-informed" approach, which means they taught the AI the laws of fluid flow and nutrient transport. When they tested this AI against the real photos, it was a champion. In predicting how the bacteria would look 24 minutes after the training data ended (a "farther-ahead extrapolation"), the AI got it right 93% of the time ( $R^2 = 0.93$ ). Compare that to the old-school math model, which only got 85% right, and a purely data-memorizing model, which got 86% right. The AI was the only one that could reliably predict the future behavior of the bacteria without getting confused.

But the team didn't stop at just predicting the future; they wanted to understand the "why." They used their AI to work backward, a process called "inverse closure learning." Imagine trying to figure out the secret recipe of a cake just by tasting the final product. The AI looked at the messy, real-world growth patterns and figured out a new, more accurate "self-inhibition" rule that the bacteria were actually following. This new rule fit the data better than the old textbook rules, suggesting that the bacteria's growth is tightly regulated by how crowded they get and how much food is left, in a way that standard models had missed.

Finally, the researchers tested if their AI could learn from one situation and apply that knowledge to a new one with very little data. This is called "transfer learning." They took the AI that had studied the 69-minute growth and asked it to predict a different, shorter growth experiment where they only had 20% of the usual data. Without any help, a new AI trained from scratch on this tiny amount of data failed miserably, getting only 63% right. But the pre-trained AI, which had already learned the "rules of the game" from the first experiment, adapted quickly and got it right 91% of the time.

In short, this paper suggests that by combining real-world observations with the unbreakable laws of physics, we can build smarter computer models that not only predict how microscopic bacterial cities grow but also help us understand the hidden rules they follow. It's a step toward better managing biofilms in everything from water treatment to medical devices, proving that sometimes, the best way to understand the tiny world is to teach a computer to respect both the data and the laws of nature.

Technical Summary: A Physics-Informed Computational Framework for Forward Prediction and Inverse Closure Learning of Pore-Scale Biofilm Growth

Problem Statement
Modeling biofilm growth in porous media presents a significant multiphysics challenge due to the complex, spatiotemporal coupling of fluid flow, nutrient transport, and microbial kinetics. Traditional continuum models often rely on simplified physical assumptions that fail to capture pore-scale heterogeneities, leading to parameter uncertainty and structural errors, particularly regarding closure relationships for biofilm expansion. Conversely, purely data-driven approaches struggle with generalization and physical consistency when extrapolating beyond training data. Furthermore, directly observing pore-scale dynamics in opaque porous media is difficult, necessitating the use of microfluidic platforms which, while transparent, generate complex data requiring robust computational frameworks to interpret. The core challenge addressed is how to accurately predict biofilm evolution and identify effective observation-scale closure laws (specifically self-inhibition mechanisms) by integrating experimental microfluidic data with governing physical equations.

Methodology
The study proposes a physics-informed computational framework integrating a modified flow–transport–growth model with Physics-Informed Neural Networks (PINNs).

Experimental Setup: Time-resolved microfluidic imaging was conducted using a chip colonized with Bacillus subtilis and fed a glucose solution. High-dynamic-range imaging captured biofilm growth over 24 hours. Data processing converted transmitted light intensity into spatially resolved fields of biomass ( $M$ ), porosity ( $n$ ), and permeability ( $k$ ).
Physical Modeling:
- Flow: The Darcy–Brinkman–Stokes (DBS) equation was used to uniformly describe flow in both free channels and porous biofilm regions, avoiding complex remeshing.
- Nutrient Transport: A Convection–Diffusion–Reaction (CDR) equation modeled substrate transport and consumption based on Monod kinetics.
- Biomass Evolution: A modified Diffusion–Reaction (DR) equation was formulated. Crucially, the authors introduced a density-dependent growth formulation incorporating a self-inhibition factor ( $S$ ). This factor captures a non-monotonic (S-shaped) relationship between biomass and expansion, transitioning from substrate-limited growth at low biomass to self-inhibition (carrying capacity) at high biomass. The biofilm diffusion coefficient ( $D_b$ ) was also made dependent on substrate concentration and the self-inhibition factor.
Numerical Calibration: A baseline coupled multiphysics model was implemented in COMSOL Multiphysics. Parameters were calibrated using a sequential quadratic programming (SQP) algorithm to minimize the mismatch between simulation and experimental biomass fields (using MSE and SSIM metrics).
PINN Architecture: A deep neural network was designed with an input mapping layer, a residual connection module (featuring parallel sub-networks of varying depths), and feature fusion layers. The network takes spatiotemporal coordinates as input and outputs velocity, pressure, substrate concentration, and biomass.
- Training: The loss function combined data-fitting loss (Huber loss against experimental biomass) with residual losses from the governing PDEs (DBS, CDR, DR) and boundary/initial conditions.
- Inverse Closure Learning: A fully connected neural network (FCNN) was embedded within the PINN to learn an effective self-inhibition correction factor ( $S'$ ) and a diffusion coefficient parameter ( $b_{inv}$ ) directly from data, replacing the fixed functional form of $S$ .
- Transfer Learning: Pre-trained NN-PINN models were adapted to a new kinetic scenario with sparse data by fine-tuning network parameters, comparing strategies of freezing different network layers.

Key Contributions

Modified Biofilm Model: The study proposes a density-dependent biofilm diffusion model with an explicit self-inhibition term, derived from the observation of non-monotonic biomass-expansion dynamics in microfluidic experiments.
PINN Framework for Biofilms: It establishes a PINN framework specifically for pore-scale biofilm growth, enabling joint forward prediction and inverse inference under coupled PDE constraints.
Inverse Closure Learning: The framework demonstrates the ability to inversely learn an effective observation-scale closure (the self-inhibition factor) that refines the physical model's structural accuracy without requiring explicit molecular mechanisms.
Data Efficiency via Transfer Learning: The study shows that a pre-trained PINN can serve as a reusable prior, allowing accurate adaptation to related scenarios with significantly reduced labeled data (sparse supervision).

Results

Experimental Analysis: Imaging revealed that biofilm growth rate is highest near free channels (nutrient-rich) and decreases in the interior. A clear S-shaped relationship was observed between biomass and growth rate, confirming a transition from substrate limitation to self-inhibition.
Forward Prediction: In extrapolation tests (predicting 48–69 minutes based on 0–45 minutes training data), the PINN achieved an $R^2$ of 0.93. This outperformed the calibrated numerical model ( $R^2 = 0.85$ ) and the purely data-driven model ( $R^2 = 0.86$ ). The PINN maintained accuracy in farther-ahead extrapolation where the data-driven model degraded significantly due to a lack of physical constraints.
Inverse Closure Learning: The NN-PINN successfully learned a correction factor ( $S'$ ) that exhibited a trend similar to the original $S$ but resulted in smaller residuals for the diffusion-reaction equation. This indicated that the data-guided closure better aligned with the governing physics and reduced structural mismatch. Even with reduced data constraints (removing boundary and data-fitting losses), the NN-PINN remained more stable than the standard PINN.
Transfer Learning: When applied to a new scenario with only 20% of the original training data, the transferred NN-PINN achieved an $R^2$ of 0.91, capturing fine-scale biomass patterns. In contrast, a model trained from scratch on the sparse data achieved only $R^2 = 0.63$ . Strategies that did not freeze layers or froze only shallow layers performed best.

Significance
The paper claims that this framework provides a data-efficient computational strategy for predictive simulation and inverse closure learning in coupled biofilm-mediated transport systems. By integrating microfluidic observations with mechanistic constraints, the approach addresses the limitations of traditional models (structural errors) and pure data-driven models (poor generalization). The ability to learn effective observation-scale closures and adapt pre-trained models to sparse data scenarios suggests a pathway for more robust modeling of biofilm processes in engineered porous media, such as bioremediation and microbial electrochemical technologies. The authors note that future work will extend the framework to include variable boundary conditions and explicit mechanical responses like deformation and detachment.

More like this