⚡ electrical engineering

Physics-Informed Machine Learning for Radiation-Free Water Liquid Ratio Estimation Using Coriolis Flow Measurements

This study introduces a radiation-free, physics-informed machine learning framework that accurately estimates the Water Liquid Ratio in three-phase oil–gas–water flows using only Coriolis mass flow measurements and a novel Harmonic Mixture Density Mean Factor (HMDMF), achieving high predictive performance (R² = 0.8943) and offering a cost-effective alternative to traditional gamma-based systems.

Original authors: Hamza Zidoum, Majid Al Yaroubi

Published 2026-07-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hamza Zidoum, Majid Al Yaroubi

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 you are trying to figure out exactly how much water is mixed into a bucket of oil and gas flowing through a pipe. This is a huge deal for oil companies because they need to know the exact recipe of what they are pumping out of the ground to manage their business.

Traditionally, to solve this "liquid recipe" puzzle, companies use a special, expensive, and slightly scary tool called a Dual Gamma Meter. Think of this like a high-tech, radiation-powered X-ray machine that looks through the pipe to count the oil, water, and gas particles. It works great, but it's costly, requires special safety permits, and involves handling radioactive materials.

The Problem:
The oil companies in Oman wanted a cheaper, safer way to do this without using radiation. They already had a different tool installed: a Coriolis Flow Meter. You can think of this meter like a very sensitive, vibrating tuning fork. As liquid flows through it, the tube wiggles. By measuring how it wiggles, the meter can tell you the total weight of the liquid and how dense it is.

However, there's a catch. If there is even a tiny bit of gas bubbles mixed in with the liquid, the vibrating tube gets confused. It's like trying to weigh a bag of feathers while someone is shaking the bag; the reading becomes inaccurate. The meter sees the "lightness" of the gas and thinks the whole liquid is lighter than it really is, making it impossible to tell how much water is actually there.

The Solution: A "Physics-Smart" AI
The researchers in this paper built a new system that uses Machine Learning (AI) to fix the Coriolis meter's confusion, but with a special twist. Instead of just feeding the AI raw numbers, they taught it a specific "physics rule" to help it understand what's happening.

Here is the core of their invention, explained simply:

The "Gas Detector" Clue (HMDMF):
The researchers created a new mathematical clue they call the Harmonic Mixture Density Mean Factor (a fancy name for a "Gas Sensitivity Score").
- The Analogy: Imagine you have a smoothie with heavy strawberries (oil), heavy water (water), and light air bubbles (gas). If you just weigh the smoothie, the air bubbles make it seem lighter.
- The researchers realized that if you do a specific type of math calculation (called a "harmonic mean") on the known weights of oil, water, and gas, the result is extremely sensitive to the presence of those air bubbles.
- They created a ratio: How heavy the mixture looks vs. how heavy the math says it should be if there were no gas.
- This ratio acts like a secret signal. Even without a gas sensor, this number tells the AI: "Hey, there are bubbles messing up the weight reading!"
The Training:
They took data from 735,034 real-world measurements from 26 oil wells in Oman. They fed this data into five different types of AI "students" (including Random Forests, which are like a committee of decision-makers, and Neural Networks, which are like digital brains).
- Test 1 (The Blind Student): They gave the AI only the raw weight and flow speed.
  - Result: The AI was terrible. It couldn't guess the water amount at all. It was like trying to guess the ingredients of a soup just by looking at the pot's handle.
- Test 2 (The Smart Student): They gave the AI the raw data plus the new "Gas Sensitivity Score" (the HMDMF).
  - Result: The AI suddenly became a genius. It could accurately predict the water content with about 89% accuracy.

The Big Takeaway:
The study proves that you don't need the expensive, radioactive X-ray machine anymore. By using a standard vibrating meter (Coriolis) and teaching the AI a specific physics-based "trick" (the HMDMF score) to spot gas bubbles, they can figure out the water-to-oil ratio just as well as the high-tech radiation method.

In Summary:

Old Way: Use a radioactive X-ray to see inside the pipe. (Expensive, risky).
New Way: Use a vibrating meter + a smart math trick + an AI brain. (Cheaper, safer, and just as accurate).

The researchers found that the "math trick" (the HMDMF) was the most important ingredient. Without it, the AI failed. With it, the AI succeeded, proving that understanding the physics of the problem is the key to making the computer smart enough to solve it.

Technical Summary: Physics-Informed Machine Learning for Radiation-Free Water Liquid Ratio Estimation

Problem Statement
Accurate determination of the Water Liquid Ratio (WLR) and pure liquid density in three-phase (oil–water–gas) flow systems is critical for reservoir management and production optimization. Conventional solutions rely on Multiphase Flow Meters (MPFMs) integrating dual gamma-ray sensors for phase-fraction resolution. While accurate, these systems present significant drawbacks: high capital and maintenance costs, strict regulatory compliance requirements, radiation safety hazards, and calibration complexity. Consequently, operators in Oman and the GCC region are seeking radiation-free alternatives.

Existing Low-Performance Testers (LPT) utilize Coriolis Mass Flow Meters (CMFM) to measure liquid flow rate and apparent mixture density. However, in the absence of a dedicated gas separator, residual gas entrainment in the liquid line creates a "semi-three-phase" condition. This entrainment depresses the apparent Coriolis density below the true liquid value, introducing significant uncertainty when calculating WLR using standard volumetric mixing equations. While machine learning (ML) has been applied to correct Coriolis errors in two-phase flows or using gamma-ray inputs, no validated framework exists for estimating WLR and pure liquid density using only standard Coriolis field outputs (without gamma instrumentation or internal diagnostics) on real-world production data.

Methodology
The study proposes a physics-informed machine learning framework trained on a dataset of 735,034 one-minute records collected from 26 oil wells across three LPT units at the Daleel Petroleum field in Oman (January–August 2025).

Data Preprocessing: The pipeline involved PVT normalization to convert laboratory densities to in-situ conditions. Rigorous filtering removed missing values, instrument faults (e.g., WLR > 100%), and outliers. A distribution-free, quantile-based stability analysis established a minimum reliable flow rate threshold of 3.0 m³/d to ensure signal-to-noise ratio integrity.
Feature Engineering (The Harmonic Mixture Density Mean Factor - $k_H$ ): The core innovation is the derivation of $k_H$ $k_{H}$ , a dimensionless, gas-sensitive index.
- The apparent mixture density ( $\rho_{mix}$ ) is the arithmetic mean of phase densities weighted by volume fractions.
- The harmonic mean density ( $\rho_{HMDM}$ ) is calculated based on the inverse of the phase densities.
- $k_H$ is defined as the ratio $\rho_{mix} / \rho_{HMDM}$ .
- Analytically, $k_H$ is shown to decrease monotonically as the Gas Volume Fraction (GVF) increases. Unlike generic engineered features, $k_H$ is derived from three-phase mixing equations to encode the disproportionate influence of entrained gas on the harmonic mean, providing an implicit signal of GVF without requiring direct gas measurement.
Experimental Design: Five machine learning models (Linear Regression, Support Vector Regression, Random Forest, Gradient Boosting, and a custom Physics-ANN) were evaluated in a two-step protocol:
- Step 1 (Baseline): Trained on raw inputs only: Liquid Flow Rate ( $Q_l$ ) and Gas Density ( $\rho_g$ ).
- Step 2 (Full): Retrained with the addition of $k_H$ as a third input.
- This design isolates the specific contribution of $k_H$ to predictive accuracy across different model families.
Validation: Models were evaluated using an 80/20 random split and a 5-fold temporal cross-validation protocol to ensure results were not artifacts of temporal leakage (where consecutive time-stamped records share identical conditions).

Key Results

Baseline Performance: Without $k_H$ , all five models performed poorly, with $R^2$ values ranging from 0.083 to 0.194. This confirms that raw Coriolis measurements of flow rate and gas density alone are insufficient for WLR estimation.
Impact of $k_H$ : Incorporating $k_H$ resulted in a dramatic, uniform improvement across all model architectures. $R^2$ values increased by +0.70 to +0.79 across the board.
Best Performing Model: The Random Forest model achieved the highest accuracy with $R^2 = 0.8943$ and RMSE = 0.0854 on test data. Gradient Boosting ( $R^2 = 0.8940$ ) and SVR ( $R^2 = 0.8920$ ) followed closely. The ANN achieved $R^2 = 0.8825$ .
Temporal Validation: The temporal cross-validation yielded a mean $R^2$ of 0.8913, nearly identical to the random-split results. This confirms the model's generalization capability and that the performance is not due to temporal data leakage.
Flow Thresholds: Filtering for higher flow rates (10.0 m³/d) improved performance slightly (RMSE reduced by ~12.8%), but aggressive filtering (24.0 m³/d) yielded diminishing returns, indicating a trade-off between data quality and quantity.

Significance and Claims
The paper claims three primary contributions:

The $k_H$ Index: It introduces $k_H$ as a novel, physics-derived surrogate observable that quantifies gas entrainment effects. The study demonstrates that $k_H$ is the primary driver of predictive accuracy, providing a computable signal of GVF without additional sensors.
Model-Agnostic Improvement: The performance gain is consistent across linear, kernel-based, ensemble, and neural network architectures. This proves that the improvement stems from the physical information content of the feature rather than the inductive bias of a specific algorithm.
Radiation-Free Alternative: The framework establishes a viable, cost-effective method for estimating WLR and pure liquid density using only standard Coriolis outputs and laboratory phase densities, eliminating the need for dual gamma meters.

The authors conclude that while the Random Forest model with $k_H$ offers a robust solution for well-test measurements, the remaining performance gap (relative to unity) is likely due to the absence of internal Coriolis vibrational diagnostics (e.g., damping, drive gain) and the lack of temporal sequence modeling, rather than a deficiency in the $k_H$ feature itself. The study positions this work as a significant step toward radiation-free virtual flow metering in petroleum production.

More like this