Compartment Modelling of Multiphase Reactors using Unsupervised Clustering

This paper introduces CLARA, a software toolbox that utilizes unsupervised clustering of CFD data to automatically generate accurate and computationally efficient compartment models for the real-time control and optimization of multiphase reactors.

The Gist

The Big Problem: The "Super-Computer" vs. The "Real-Time" Need

Imagine you are trying to design a giant industrial mixer, like a massive vat used to make medicine or chemicals. Inside this vat, bubbles rise, liquids swirl, and chemical reactions happen. To understand exactly what is happening, scientists use a super-powerful simulation called CFD (Computational Fluid Dynamics).

Think of CFD as a high-definition, 4K movie of the inside of the vat. It shows every single bubble, every swirl of liquid, and every tiny change in concentration. It is incredibly accurate, but it is also incredibly slow. Running one of these simulations can take days or even weeks on a supercomputer.

The Problem: You can't use a slow, 4K movie to control a machine in real-time. If you want to adjust the mixer right now to prevent a reaction from going wrong, or to design a new vat efficiently, you need a fast, lightweight sketch that still captures the important details.

The Solution: CLARA (The "Smart Partitioner")

The authors introduce a new software tool called CLARA. Its job is to turn that slow, heavy 4K movie into a fast, simple sketch called a Compartment Model (CM).

Instead of tracking every single molecule, CLARA divides the giant vat into a few distinct "rooms" or compartments. Inside each room, everything is perfectly mixed (like a cup of coffee you've stirred well). The model only needs to track the average concentration in each room and how much liquid flows between the rooms.

The Analogy:

• CFD is like counting every single grain of sand on a beach to understand the tide.
• CLARA is like dividing the beach into 10 large buckets, measuring the average wetness of the sand in each bucket, and tracking how water moves between the buckets.

How CLARA Works (The Magic Trick)

The paper explains that CLARA doesn't just guess where to draw the lines between these "rooms." It uses Unsupervised Clustering, which is a type of Artificial Intelligence (AI) that finds patterns on its own.

1. The Input: CLARA looks at the data from the slow CFD simulation. It sees where the liquid is moving fast, where it is slow, and where the chemical concentration is high or low.
2. The Grouping: It groups the tiny cells of the simulation together based on what they have in common.
   • Analogy: Imagine a classroom of students. Instead of listing every student individually, the teacher groups them by "who sits near whom" and "who has the same homework." CLARA does this with fluid cells.
3. The Rules: The paper highlights two main ways CLARA groups these cells:
   • K-Means: Tries to make the groups round and compact (like grouping students by who is closest to the center of the room).
   • Hierarchical Clustering: Builds groups by merging neighbors, ensuring that the "rooms" are physically connected (like grouping students by who is sitting in the same row).
4. The Safety Check: A major innovation in this paper is a "mass conservation" check. Sometimes, when you simplify a complex system, you accidentally create or destroy fluid (like a leaky bucket). CLARA has a built-in "plumber" that adjusts the flow rates between rooms to ensure that what goes in equals what comes out, keeping the math physically correct.

The Test: The "Quarter-Gassed" Bubble Column

To prove it works, the authors tested CLARA on a specific, tricky scenario: a Bubble Column Reactor.

• The Setup: Imagine a tall tank where gas is injected only from the right side of the bottom. This creates a chaotic situation: the right side is bubbling and mixing, while the left side is quiet and stagnant.
• The Challenge: They added a chemical reaction that eats up oxygen. They tested three types of reactions:
  • 1st and 2nd Order: These reactions are "easy." They stop quickly if oxygen runs out, so the whole tank stays fairly uniform.
  • 0.5th Order: This is the "hard" test. This reaction keeps going even when oxygen is very low. This creates a massive difference between the bubbly right side and the starving left side.

What They Found

1. Accuracy: CLARA was able to recreate the complex chemical patterns of the slow CFD simulation very accurately, but much faster.
2. The "Feature" Secret: The most important finding was about what data CLARA uses to group the rooms.
   • If you tell CLARA to group cells based on flow speed or bubble size, it fails to capture the chemical differences.
   • If you tell CLARA to group cells based on chemical concentration, it works much better.
3. The "Too Many Rooms" Trap: The paper discovered a counter-intuitive result. You might think, "If I make 50 rooms instead of 5, it will be more accurate."
   • Surprise: For this specific type of reaction, making too many rooms actually made the model worse.
   • Why? When you cut the tank into too many thin slices, you accidentally slice through areas where the liquid is naturally mixing due to turbulence (chaos). The simplified model can't see this "invisible mixing," so it creates fake chemical gradients.
   • The Sweet Spot: They found that using a moderate number of rooms (around 5 to 10) was the perfect balance. It captured the big differences without breaking the natural mixing.

The Conclusion

The paper concludes that CLARA is a powerful, open-source toolbox that can automatically turn slow, complex fluid simulations into fast, simple models.

• It handles multiphase flows (gas and liquid together), which previous tools struggled with.
• It ensures mass is conserved (no leaks).
• It proves that for complex chemical reactions, you don't need a million tiny rooms; you just need the right number of rooms grouped by the right chemical features.

This tool allows engineers to design better reactors and control them in real-time without needing a supercomputer to wait days for an answer.

Technical Summary

1. Problem Statement

Industrial-scale multiphase reactors (e.g., bioreactors, gas-liquid reactors) exhibit significant spatial heterogeneities in concentration, temperature, and pH due to non-ideal mixing. These gradients critically impact reaction rates, yield, and process stability, particularly for non-linear kinetics (e.g., fractional-order reactions in oxygen-limited bioprocesses).

• Limitation of Full CFD: While Computational Fluid Dynamics (CFD) provides high-resolution insights, it is computationally prohibitive for real-time control, long-duration simulations (e.g., fed-batch fermentations), and Model Predictive Control (MPC).
• Limitation of Existing Reduced Order Models (ROMs): Traditional Compartment Models (CM) or Chemical Reactor Networks (CRN) often rely on manual zoning or fixed grids. Existing automated methods using unsupervised learning are primarily limited to single-phase flows, lack robust handling of interphase mass transfer, and often suffer from mass conservation errors due to interpolation. Furthermore, no open-source tool exists that is solver-agnostic (compatible with OpenFOAM, ANSYS Fluent, LBM) and capable of handling multiphase flows.

2. Methodology: The CLARA Toolbox

The authors introduce CLARA (Clustering Algorithm and Applications), an open-source software toolbox that automates the generation of Compartment Models from time-averaged CFD data using unsupervised machine learning.

Core Workflow

1. Data Input: Accepts time-averaged CFD data (velocity, volume fraction, concentration, turbulence) from various solvers (OpenFOAM, ANSYS Fluent, SimVantage LBM).
2. Unsupervised Clustering: Partitions the reactor geometry into spatially connected zones (compartments) based on user-selected features.
   • Algorithms: Implements K-Means (with a post-processing graph reassignment step to enforce spatial connectivity) and Hierarchical Agglomerative Clustering (which inherently respects spatial adjacency).
3. Mass Conservation Optimization: Addresses numerical inconsistencies in CFD data by solving a constrained optimization problem (using SLSQP) to adjust inter-compartment flow rates. This ensures strict steady-state mass balance without creating unphysical flow paths between disconnected regions.
4. Multiphase Modeling:
   • Treats each phase (e.g., gas and liquid) within the same spatial cluster boundaries but with distinct volumes ($V_{gas} = V_{tot} \cdot \alpha_{gas}$).
   • Models Interphase Mass Transfer explicitly within each compartment using a $k_L a$ coefficient and Henry's law, ensuring mass transfer occurs only between corresponding gas-liquid compartment pairs.
5. Simulation: Solves the resulting system of Ordinary Differential Equations (ODEs) for species concentration.

Mathematical Framework

The concentration balance for each compartment $i$ is defined as:
$$\frac{dc_i}{dt} = s_{conv} + s_{react} + s_{MT}$$
Where $s_{conv}$ accounts for convective transport derived from the CFD flow matrix, $s_{react}$ handles non-linear reaction kinetics, and $s_{MT}$ models mass transfer between phases.

3. Key Contributions

• First Open-Source Multiphase CM Generator: CLARA is the first toolbox to automate compartment generation for multiphase flows using unsupervised learning, supporting multiple CFD solvers.
• Robust Mass Conservation: Introduces a two-stage correction (boundary scaling and local flow optimization) to guarantee mass conservation, a critical requirement often missed in automated clustering.
• Explicit Mass Transfer: Unlike previous clustering studies, CLARA explicitly models interphase mass transfer within the generated network, validated for gas-liquid systems.
• Systematic Feature Analysis: Provides a rigorous investigation into how clustering features (concentration, volume fraction, turbulent kinetic energy) and algorithms affect model accuracy in reactive multiphase systems.

4. Results and Validation

Analytical Verification

The framework was verified against analytical solutions for single-phase (constant, linear, quadratic velocity profiles) and two-phase mass transfer cases.

• Finding: The relative error in reactor yield decreases monotonically as the number of compartments ($n_c$) increases, confirming the mathematical correctness of the clustering and mass balance procedures.

Case Study: Quarter-Gassed Bubble Column

A reactive gas-liquid bubble column (Oxygen dissolving in water) with non-linear reaction kinetics (0.5th, 1st, and 2nd order) was simulated.

• Hydrodynamics: The system exhibited strong asymmetry with a gas plume on one side and a recirculation loop on the other, creating steep concentration gradients.
• Reaction Orders:
  • 1st & 2nd Order: Resulted in relatively homogeneous concentration fields. CMs achieved low errors (<10%) regardless of the clustering strategy.
  • 0.5th Order: Produced highly heterogeneous concentration fields (spanning orders of magnitude), serving as the most demanding benchmark.
• Feature Selection Impact:
  • Flow-only features (e.g., volume fraction $\alpha_{liq}$) failed to capture concentration gradients, yielding errors higher than a single-compartment model.
  • Concentration-based features ($c_{O_2}$) achieved the lowest error ($\approx 4.5\%$) at low cluster counts ($n_c = 5$) using Agglomerative clustering.
• The "Error Increase" Phenomenon: Counter-intuitively, increasing $n_c$ beyond a certain point (e.g., $>7$) increased the error for the 0.5th-order case when using only concentration features.
  • Cause 1 (Geometric Fragmentation): High $n_c$ creates thin, iso-concentration bands that bridge physically distinct zones (gas-rich plume vs. gas-free return flow). This artificially applies mass transfer to oxygen-depleted zones.
  • Cause 2 (Suppression of Turbulent Mixing): Compartment boundaries in the CM only allow convective exchange. As $n_c$ increases, boundaries bisect well-mixed turbulent zones. Since the CM lacks a sub-grid turbulent mixing term, this truncation creates artificial gradients.
  • Verification: A simulation with suppressed turbulent diffusivity ($S_{ct} = 10^7$) showed that error decreased monotonically with $n_c$, confirming that turbulent mixing truncation is the primary cause of the error increase in the baseline case.
• Algorithm Comparison: Hierarchical Agglomerative Clustering outperformed K-Means for this complex, poorly mixed system. Agglomerative clustering naturally produced spatially coherent, concentric shells, whereas K-Means (even with graph reassignment) produced fragmented, geometrically irregular compartments.

5. Significance and Conclusion

• Practical Guidelines: The study identifies a "sweet spot" of 3 to 10 compartments for such systems. This range balances spatial resolution with the validity of the well-mixed assumption, avoiding the pitfalls of over-clustering in turbulent flows.
• Engineering Relevance: While local concentration fields may have errors, CLARA predicts the total reactor reaction rate with <2% accuracy. This is because the system is mass-transfer limited, and the integral quantity is well-preserved even if local gradients are imperfectly resolved.
• Impact: CLARA enables the creation of computationally efficient surrogates (speedup factors >2000x compared to CFD) suitable for Digital Twins and Real-time Model Predictive Control (MPC) of complex multiphase reactors.
• Future Work: The authors suggest incorporating physics-informed features (e.g., residence time, local Hatta number) and developing methods to model turbulent transport across compartment boundaries to further improve accuracy at higher resolutions.

In summary, CLARA provides a robust, automated, and mathematically verified framework for transforming high-fidelity multiphase CFD data into reduced-order compartment models, bridging the gap between detailed simulation and real-time industrial control.