Data-driven Learning of Probabilistic Model of Binary Droplet Collision for Spray Simulation

This paper presents a novel, data-driven probabilistic model for binary droplet collisions, developed using LightGBM on a comprehensive experimental dataset and translated into a multinomial logistic regression framework to enable accurate, stochastic implementation in spray simulations.

The Gist

Imagine you are watching two raindrops collide in mid-air. Sometimes they merge into a bigger drop. Sometimes they bounce off each other like tiny rubber balls. Sometimes they smash apart into a mist of smaller droplets.

For decades, scientists have tried to write "rules" to predict exactly what will happen. But nature is messy. If you repeat the exact same experiment twice, the drops might do something different the second time, especially when they are on the edge between merging and breaking. Traditional computer models treat these rules like rigid walls: "If the speed is X, the result is Y." But in reality, the boundaries are fuzzy, like a foggy horizon rather than a sharp fence.

This paper introduces a new, smarter way to predict these collisions using Machine Learning. Here is the story of how they did it, explained simply:

1. The Problem: The "Rulebook" Was Too Rigid

Think of traditional models as a strict traffic cop. They say, "If you drive at 30 mph, you must stop." But in the real world, sometimes a driver at 30 mph might speed up, or sometimes they might stop early. The old models couldn't handle this "maybe." They also missed many rare types of collisions because the data they learned from was incomplete.

2. The Solution: A "Super-Student" (LightGBM)

The researchers gathered a massive library of 33,540 real-life collision experiments from 26 different studies. They fed this data into a powerful AI algorithm called LightGBM.

• The Analogy: Imagine a super-student who has watched 33,000 hours of raindrop collision videos. Instead of memorizing a rigid rulebook, this student learns the patterns and the feel of the collisions.
• The Result: This AI became incredibly good at guessing the outcome. It got the answer right 99.2% of the time. More importantly, it learned that in the "foggy" transition zones, the answer isn't just "A" or "B," but "There's a 60% chance of A and a 40% chance of B."

3. The Translation: From "Black Box" to "Clear Recipe"

AI models are often "black boxes"—you put data in, and an answer comes out, but you don't know how the AI decided. For engineers building spray simulations (like for car engines or inkjet printers), they need to know the "why."

• The Analogy: The AI is like a genius chef who makes a perfect dish but won't tell you the recipe. The researchers took the chef's intuition and translated it into a simple, written recipe (using a method called Multinomial Logistic Regression).
• The Result: They turned the complex AI brain into a set of mathematical equations that anyone can read. This "recipe" kept 93.2% of the AI's accuracy but made the logic transparent and easy to use in computer simulations.

4. The Final Step: The "Biased Dice"

Now, the computer has the probabilities (e.g., "60% chance of merging, 40% chance of bouncing"). But a computer simulation needs to make a single, definite choice for every single drop collision it calculates.

• The Analogy: Imagine you have a weighted, 8-sided die. Each side represents a different collision outcome. The die is "biased," meaning the sides with higher probabilities are heavier and more likely to land face-up.
• The Process: When the simulation needs to decide what happens to a drop, it "rolls the dice" based on the probabilities calculated by the model.
  • If the model says "99% chance of merging," the die is almost guaranteed to land on "merge."
  • If the model says "50/50 chance," the die is fair, and the outcome is truly random, just like in real life.

Why Does This Matter?

This approach is a game-changer for spray simulations. Whether it's designing a fuel injector for a rocket, creating a perfect spray of medicine, or predicting how rain forms in clouds, engineers need to know how droplets behave.

• Old Way: "It will definitely merge." (Often wrong in tricky situations).
• New Way: "It's a toss-up, so let's roll the dice and see what happens." (Physically accurate and realistic).

Summary

The authors built a digital crystal ball for droplet collisions. They trained a smart AI on thousands of experiments, translated its "gut feeling" into a clear mathematical recipe, and gave it a "biased dice" to make realistic, random decisions. This allows scientists to simulate sprays with a level of realism and uncertainty that was previously impossible, bridging the gap between messy real-world physics and clean computer code.

Technical Summary

1. Problem Statement

Binary droplet collisions are fundamental to processes like fuel atomization, inkjet printing, and cloud formation. Traditional deterministic models used in spray simulations rely on analytical or empirical correlations to define sharp boundaries between collision regimes (e.g., coalescence, bouncing, separation). However, these models face significant limitations:

• Inability to Capture Stochasticity: They fail to represent the transitional and stochastic behaviors observed in experiments, where identical conditions can yield different outcomes.
• Limited Regime Coverage: Existing models often cover a reduced number of regimes (typically 3–4) and narrow parameter ranges, neglecting complex outcomes like finger separation or splashing.
• Data Sparsity and Uncertainty: Experimental data is unevenly distributed (e.g., scarce data for low Weber numbers), and previous models often rely on extrapolation or ignore experimental uncertainties.
• Black-Box vs. Interpretability: While machine learning (ML) offers high accuracy, many approaches are "black boxes" lacking physical interpretability or explicit analytical forms required for integration into Computational Fluid Dynamics (CFD) codes.

2. Methodology

The authors propose a hybrid, three-stage data-driven framework that combines high-accuracy machine learning with explicit probabilistic modeling to generate a simulation-ready tool.

A. Data Collection and Feature Engineering

• Dataset: A comprehensive database of 33,540 experimental data points was compiled from 26 previous studies.
• Regimes: The data covers eight distinct collision regimes:
  1. Soft coalescence
  2. Bouncing
  3. Hard coalescence
  4. Reflexive separation
  5. Stretching separation
  6. Rotational separation
  7. Finger separation
  8. Splashing
• Input Parameters: Five dimensionless parameters were identified via Buckingham–Pi analysis:
  • Weber number ($We$): 0–2000
  • Ohnesorge number ($Oh$):$2.7 \times 10^{-3}$–$5.5 \times 10^{-1}$
  • Impact parameter ($B$): 0–1
  • Size ratio ($\Delta$): 1–5
  • Ambient pressure ratio ($P$): 0.6–9.0

B. Stage 1: LightGBM Probabilistic Classification

• Algorithm: The Light Gradient-Boosting Machine (LightGBM) was employed to learn the complex, non-linear relationships between input parameters and collision outcomes.
• Key Features:
  • Utilizes Gradient-based One-Side Sampling (GOSS) to focus on informative samples.
  • Uses Exclusive Feature Bundling (EFB) to handle sparse features efficiently.
  • Outputs class probabilities rather than hard labels, naturally capturing the "fuzzy" nature of regime transitions.
• Performance: Achieved a macro-averaged accuracy of 99.2% with a cross-validation F1-score of 0.921.

C. Stage 2: Multinomial Logistic Regression (Analytical Projection)

• Objective: To convert the implicit "black-box" tree predictions of LightGBM into an explicit, interpretable analytical form suitable for CFD integration.
• Process: The probability fields generated by LightGBM were projected onto a multinomial logistic regression model.
  • Input features were expanded using a second-degree polynomial basis (including interaction terms like $We \times B$).
  • The model minimizes cross-entropy loss to fit the probabilities.
• Result: This step preserves 93.2% accuracy while providing explicit mathematical expressions for regime probabilities, enabling physical interpretation and visualization.

D. Stage 3: Biased-Dice Sampling (Stochastic Realization)

• Mechanism: To generate definite outcomes for individual droplet collisions in a simulation, the continuous probabilities from the logistic regression are converted into discrete outcomes using a biased-dice sampling mechanism (multinomial distribution sampling).
• Significance: This approach avoids artificial determinism. Instead of always selecting the highest probability class, it samples based on the probability distribution, thereby reproducing the inherent stochasticity and variability observed in physical experiments, especially in transitional regions.

3. Key Results

• High Accuracy: The LightGBM classifier achieved 99.2% accuracy, while the projected logistic regression model maintained 93.2% accuracy, demonstrating that the analytical form successfully distills the complex tree-based logic.
• Fuzzy Boundaries: The model successfully maps continuous inter-regime transitions. Visualization of the $We-B$ space shows smooth probability contours rather than sharp lines, accurately reflecting the physical uncertainty near regime boundaries (e.g., between bouncing and coalescence).
• Robustness: The stochastic classification (biased-dice) showed high stability across 30 independent realizations, with mean accuracy >0.94 and low standard deviations, confirming that the method preserves regime integrity while introducing necessary variability.
• Comprehensive Coverage: The model is the first to simultaneously handle eight collision regimes across a wide range of five dimensionless parameters, including high-pressure and high-size-ratio conditions previously underrepresented in models.

4. Key Contributions

1. First Probabilistic, High-Dimensional Model: This work presents the first droplet collision model derived from experimental data that is both probabilistic and covers eight regimes across a high-dimensional parameter space.
2. Hybrid ML-Physics Framework: It bridges the gap between high-accuracy "black-box" ML (LightGBM) and interpretable physics-based modeling by projecting ML outputs into explicit logistic regression equations.
3. Stochastic Sampling Mechanism: The introduction of the biased-dice sampling allows for the generation of physically consistent, stochastic outcomes in deterministic simulation codes, addressing the long-standing issue of neglecting transitional uncertainty.
4. Simulation-Ready Integration: The final output is a user-friendly, analytical probabilistic model that can be directly integrated into Eulerian–Lagrangian spray simulation frameworks (CFD).

5. Significance

This research offers a paradigm shift in spray simulation modeling. By moving from deterministic regime maps to probabilistic, data-driven classifiers, the model significantly improves the fidelity of predicting droplet size distributions, atomization behaviors, and subsequent combustion or evaporation processes. It provides a robust solution for handling the inherent uncertainty in multiphase flows, making it a critical tool for optimizing combustion engines, pharmaceutical spray drying, and meteorological modeling. The framework is also scalable, capable of accommodating additional parameters as new experimental data becomes available.