🤖 AI

Accelerated and data-efficient flow prediction in stirred tanks via physics-informed learning

This study demonstrates that physics-informed implicit neural representations can significantly improve the accuracy and stability of flow field predictions in industrial stirred tanks, particularly in low-data regimes, by trading off increased training complexity for reduced reliance on large datasets.

Original authors: Mahdi Naderibeni, Liang Wu, David M. J. Tax

Published 2026-05-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mahdi Naderibeni, Liang Wu, David M. J. Tax

Original paper licensed under CC BY 4.0 (http://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 you are trying to predict how a giant, industrial-sized blender mixes a thick liquid. To do this accurately, scientists usually run massive computer simulations that solve complex physics equations. Think of these simulations as running a high-end video game engine in slow motion: they are incredibly accurate, but they take a long time and a lot of computer power to finish.

This paper is about finding a shortcut. The researchers wanted to train a "smart assistant" (a machine learning model) to predict how the liquid moves instantly, without having to run the heavy simulation every time. However, there's a catch: usually, these smart assistants need to watch thousands of hours of the simulation to learn the ropes. The researchers wanted to know: How much training data does this assistant actually need to be good?

Here is the breakdown of their experiment and findings, using simple analogies:

1. The Setup: The Giant Blender

The researchers studied a massive industrial tank (about the size of a small house) with four spinning paddles (impellers) and baffles (walls) to help mix things. They simulated this tank under different conditions:

How fast the paddles spin (from slow to fast).
How full the tank is (from a shallow puddle to nearly full).

They generated a "library" of 72 different simulation scenarios to use as training data.

2. The Two Types of Students

They trained two types of "students" (AI models) to learn from this library:

The Pure Data Student (MLP): This student only looks at the pictures of the liquid flow from the simulations and tries to memorize the patterns. It's like a student who only studies past exam papers without understanding the underlying math.
The Physics Student (PINN): This student also looks at the pictures, but they are also forced to follow the "rules of the universe" (the laws of physics, like conservation of mass and momentum) while they study. It's like a student who studies the exam papers and is strictly graded on whether their answers make sense according to the laws of physics.

3. The Experiment: How Much Data is Enough?

They tested these students using different amounts of training data, ranging from just 7 examples to 54 examples.

The Findings:

More Data = Better Results (But with a limit): As they gave the students more examples to study, their predictions got better. However, they hit a point of "diminishing returns." After a moderate amount of data (around 36 examples), adding more didn't help much. It's like studying for a test: reading 10 practice questions helps a lot, but reading 100 might not make you much smarter than reading 50.
The Physics Student Wins in a Crunch: When the students only had a few examples (7 or 18), the Physics Student was much better. Because they knew the rules of physics, they could guess the right answer even with very little data. The Pure Data Student got confused and made wild guesses.
The Gap Closes with More Data: As the amount of training data grew, the Pure Data Student caught up. Eventually, with enough data, both students performed almost the same. The "cheat sheet" of physics rules became less necessary because the student had seen so many examples.

4. The "Tracer" Test: Does it actually mix?

To see if the predictions were truly useful, they didn't just look at the pictures; they ran a "tracer test." Imagine dropping a drop of food coloring into the blender and watching how it spreads.

The Result: The Physics Student produced much more stable and reliable mixing patterns, especially when they had very little data. The Pure Data Student sometimes predicted that the food coloring would behave strangely or inconsistently.
The Warning: Even when the Physics Student was very accurate, trying to use their predictions inside a standard simulation software was tricky. It was like trying to fit a custom-made puzzle piece into a machine; if the piece was even slightly off in shape (even if it looked right), the machine would jam or produce weird errors.

5. The Bottom Line

You don't need a massive library: You can get very good predictions for this type of mixing tank with a surprisingly small number of training examples (perhaps fewer than 20).
Physics rules help when data is scarce: If you can't afford to run thousands of simulations to generate data, forcing the AI to follow physics laws is a game-changer. It makes the AI smarter with less information.
It's not a magic wand: While these AI models are fast and accurate, plugging them back into standard engineering software requires careful handling to avoid technical glitches.

In short, the paper shows that by teaching AI the "rules of the game" (physics), we can make it a very efficient expert that learns quickly, saving time and computing power in industrial mixing processes.

Technical Summary: Accelerated and Data-Efficient Flow Prediction in Stirred Tanks via Physics-Informed Learning

Problem Statement

Simulating fluid flows in industrial stirred vessels is computationally expensive due to the complexity of the governing partial differential equations (PDEs). While high-fidelity methods like Large Eddy Simulation (LES) and Direct Numerical Simulation (DNS) exist, even routine Reynolds-Averaged Navier-Stokes (RANS) simulations can be prohibitive for applications requiring rapid iteration, such as fermentation and crystallization process optimization. Furthermore, minor changes in operating conditions (e.g., impeller speed, liquid height) necessitate new simulations. Although Machine Learning (ML) offers a potential surrogate to accelerate these predictions, standard data-driven models typically require large training datasets, creating a trade-off between the cost of generating simulation data and predictive accuracy. This work investigates the relationship between training-set size and prediction accuracy for steady flow fields in an industrial-scale stirred vessel, specifically evaluating the efficacy of Physics-Informed Neural Networks (PINNs) in low-data regimes.

Methodology

Data Generation

The study utilizes a dataset of steady single-phase RANS simulations generated using the $k-\omega$ turbulence model (SST formulation) for a 22 $m^3$ multi-impeller Rushton turbine tank. The dataset covers a range of realistic operating conditions defined by:

Liquid height: Uniformly sampled between 1.5 m and 6.5 m.
Stirring rate: Sampled between 50 and 150 rpm.
The simulations were performed using Ansys Fluent 2024R1 with the Multiple Reference Frame (MRF) approach to capture mixer rotation. The resulting data includes spatial coordinates, velocity components, pressure, phase volume fractions, and turbulence closure parameters, sampled at 100,000 points per condition to preserve the non-uniform mesh density of the original CFD grid.

Model Architecture

The authors employ Implicit Neural Representations (INRs) to learn continuous mappings from spatial coordinates and operating parameters (stirring rate, liquid height) to flow variables.

Network Structure: A fully connected Multilayer Perceptron (MLP) with 10 hidden layers of 100 nodes.
Input Processing: To address the spectral bias of standard INRs (the tendency to learn low-frequency components first), Fourier Feature Mappings are applied to the input coordinates before they enter the MLP.
Activation: Tanh activation functions are used to ensure smooth, bounded nonlinearities suitable for physical fields.

Training Strategies and Loss Functions

Three model variants were compared:

MLP: A purely data-driven model trained exclusively on labeled CFD data using a Mean Squared Error (MSE) loss ( $L_{data}$ ).
C-MLP: Augmented with the continuity equation residual ( $R_{cont}$ ) in the loss function.
CM-MLP: Augmented with both continuity and momentum equation residuals ( $R_{cont}$ and $R_{u}$ ).

Note: Transport residuals for turbulent kinetic energy ( $k$ ) and specific dissipation rate ( $\omega$ ) were intentionally excluded from the physics-informed loss. Preliminary experiments showed that including these terms degraded training stability due to the high nonlinearity and stiffness of the SST $k-\omega$ source terms and wall-distance-dependent blending functions.

The total loss is a weighted sum of data loss and PDE residuals ( $L_{res}$ ), with the PDE term scaled by $10^{-3}$ and normalized by its detached magnitude to ensure balanced optimization.

Evaluation

Performance was assessed using:

Global Accuracy: Test MSE across all predicted variables.
Qualitative Spatial Analysis: Visual comparison of predicted fields against reference data.
Tracer Transport Simulations: Passive tracer dispersion simulations performed on "frozen" predicted flow fields to evaluate physical consistency and mixing dynamics.

Key Results

Data-Accuracy Trade-off

Monotonic Improvement: Prediction error (MSE) decreases monotonically as the training set size increases for all model classes.
Diminishing Returns: Beyond moderate dataset sizes (approximately 36–54 samples), the rate of accuracy improvement slows significantly. The performance of different model variants begins to converge, suggesting that a relatively small set of well-chosen operating points captures the dominant learnable structure of the flow.

Impact of Physics Constraints

Low-Data Regime: In scenarios with limited data (e.g., 7 or 18 samples), physics-constrained models (C-MLP and CM-MLP) significantly outperform the purely data-driven MLP. They achieve lower global MSE, reduced inter-run variability (robustness across random seeds), and more accurate spatial predictions, particularly in regions with strong velocity gradients.
High-Data Regime: As the training set grows, the relative advantage of the constraints diminishes. With sufficient data, the unconstrained MLP approaches the performance of the constrained variants.
PDE Residuals: While unconstrained MLPs reduce prediction error with more data, their PDE residuals (computed via automatic differentiation) remain relatively large. Constrained models maintain lower residuals by construction, which is critical for downstream tasks requiring field gradients.

Tracer Transport and Stability

Mixing Dynamics: Tracer simulations revealed that unconstrained models trained on small datasets exhibited higher inter-run variability and end-state bias compared to the CFD reference. The continuity constraint (C-MLP) reduced this bias and variability.
Numerical Coupling: Integrating the learned steady fields into a conventional CFD solver for transient tracer transport proved numerically delicate. Even with small global errors, local inconsistencies or discretization mismatches could introduce spurious divergence or non-physical extrema, requiring careful interpolation and consistency checks.

Significance and Contributions

The paper makes three primary contributions to the field of scientific machine learning for fluid dynamics:

Dataset Generation: The authors generated and published a controlled dataset of steady flow fields in an industrial stirred vessel using RANS ( $k-\omega$ ) simulations across a sweep of stirring rates and liquid heights.
Empirical Study of Data Efficiency: The work provides a systematic evaluation of the data–accuracy trade-off, demonstrating clear diminishing returns in accuracy beyond moderate dataset sizes and quantifying the specific benefits of physics constraints in low-data regimes.
Quantification of Physics Constraints: The study demonstrates that PDE-constrained models substantially improve accuracy and stability when data is scarce. However, it also highlights the practical challenges, including increased training complexity and the numerical fragility of coupling learned fields back into conventional CFD solvers.

The authors conclude that while physics-informed learning offers a robust pathway for data-efficient flow prediction, the benefits are most pronounced when training data is limited. As data availability increases, the marginal gain from explicit physical constraints decreases, though they remain valuable for ensuring physical consistency in derived quantities.