⚡ electrical engineering

Physics-Informed AI Surrogate Modelling of Thermal and Hydraulic Losses in Omani Crude Oil Pipelines Using RHLI-Based Analysis, Uncertainty- Quantification, and Explainable AI

This study presents a physics-informed, uncertainty-quantified AI surrogate model that utilizes a novel Rheological Heat-Loss Index and ensemble learning to rapidly and accurately predict thermal and hydraulic losses in Omani crude oil pipelines, offering a fast, interpretable alternative to traditional CFD for real-time operational optimization.

Original authors: Raja Rani Titti, Blessy Jayaron Jose, Preeti Jain

Published 2026-10-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Raja Rani Titti, Blessy Jayaron Jose, Preeti Jain

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

Moving crude oil from the ground to a refinery is a massive logistical feat, but it is also a battle against physics. As thick, heavy oil travels through long pipes, it naturally cools down, becoming even thicker and harder to push. This cooling creates a double problem: the oil loses heat to the cold desert air, and the pumps must work much harder to force the sluggish fluid through the pipe, wasting energy and money. For decades, engineers have relied on complex computer simulations to predict exactly how much heat and pressure will be lost in a specific pipeline. These simulations are incredibly accurate, but they are also slow, taking hours or even days to run a single scenario. This speed makes them useless for real-time decisions, such as adjusting a pump right now to save energy before a storm hits.

To solve this, researchers have turned to artificial intelligence, but not the kind that simply guesses based on patterns. They are building "physics-informed" models, which are computer programs trained to understand the actual laws of how fluids move and heat flows. By teaching these programs the rules of physics alongside the data, the models can make lightning-fast predictions that are just as reliable as the slow, heavy simulations. This approach is particularly vital for Oman, where the oil is heavy and the climate is harsh, and where every bit of wasted energy translates to higher costs and emissions. The goal is to create a digital tool that can tell operators exactly how much energy a pipeline is losing in a fraction of a second, allowing them to manage the flow efficiently without waiting for a slow computer to finish its work.

In a recent study, a team of researchers developed such a tool specifically for Omani crude oil pipelines. They started by creating a highly detailed virtual model of a three-segment pipeline, mimicking the real-world conditions where pipes change size and the oil flows through varying temperatures. Using advanced simulation software, they ran 962 different scenarios, changing factors like the thickness of the pipe walls, the temperature of the oil, and how fast it was moving. These simulations acted as a rigorous training ground, generating a massive dataset of "perfect" answers for how much heat, pressure, and power were lost in each specific situation. The researchers then used this data to train a new type of artificial intelligence system, teaching it to recognize the complex relationships between the oil's behavior and the energy losses.

The results were striking. The artificial intelligence models learned to predict the energy losses with near-perfect accuracy. When tested on new scenarios it had never seen before, the system could predict the amount of heat lost, the pressure drop, and the power required to pump the oil almost instantly. The most successful models were "ensemble" systems, which work by combining the predictions of many different algorithms to reach a single, highly reliable answer. For predicting heat loss, the system was so accurate that its predictions matched the slow, heavy simulations almost exactly. For the more complex task of predicting pressure loss, the system still performed exceptionally well, far outpacing older, simpler mathematical methods.

Crucially, the researchers did not just want a black box that gave answers; they wanted to understand why the answers were what they were. By using a technique that explains how the model makes decisions, they discovered that the single most important factor driving heat loss was the temperature of the air surrounding the pipe. This finding aligns perfectly with basic physics: the greater the difference between the hot oil and the cold air, the faster the heat escapes. The model also identified a new way to categorize the flow of the oil, grouping the simulations into three distinct regimes based on a custom index they created. In some conditions, the energy loss was mostly due to the oil getting cold; in others, it was mostly due to the friction of pushing the thick fluid through the pipe. This ability to distinguish between different types of energy loss helps engineers understand exactly what is happening inside the pipeline.

The study also addressed the uncertainty inherent in any prediction. In the real world, conditions are never perfectly known, so the researchers built a safety net into their model. They taught the system to provide a range of likely outcomes rather than just a single number, along with a measure of how confident it is in that range. The system proved to be extremely reliable, with its predicted ranges capturing the actual outcomes more than 97 percent of the time. This means that if an operator asks the system how much energy will be lost, the system can say, "It will be between X and Y, and we are very sure of this."

This work represents a significant step forward for the oil industry in Oman and beyond. By replacing slow, expensive computer simulations with a fast, accurate, and explainable artificial intelligence tool, pipeline operators can now monitor their systems in real time. They can adjust pumping speeds or heating systems instantly to minimize energy waste, responding to changing weather or oil conditions without delay. The researchers note that while their model was trained on simulated data, it provides a clear path toward a "digital twin" of the actual pipeline—a virtual copy that runs alongside the real one, constantly predicting performance and suggesting the most efficient way to operate. This shift from slow, reactive planning to fast, proactive management could significantly reduce the energy consumption and environmental footprint of moving crude oil across the desert.

Technical Summary: Physics-Informed AI Surrogate Modelling of Thermal and Hydraulic Losses in Omani Crude Oil Pipelines

Problem Statement
Efficient crude oil transportation is critical for the energy supply chain, particularly in Oman, where heavy and medium crude oils exhibit complex, temperature-dependent rheological behaviors. The transport of these viscoplastic fluids through pipelines incurs significant thermal and hydraulic energy losses, leading to increased pumping energy demands, operational costs, and carbon emissions. While high-fidelity Computational Fluid Dynamics (CFD) using tools like COMSOL Multiphysics can accurately model these coupled thermo-hydraulic interactions under non-isothermal, Bingham–Papanastasiou conditions, such simulations are computationally expensive. This latency renders them unsuitable for real-time operational decision-making, design optimization, or digital twin implementation. Furthermore, existing Artificial Intelligence (AI) studies in pipeline engineering often focus on single-target predictions (e.g., pressure drop only) or lack physical interpretability and uncertainty quantification.

Methodology
The study proposes a Physics-Informed AI surrogate framework designed to predict Total Heat Loss, Head Loss, and Total Power Loss for Omani crude oil pipelines. The methodology proceeds through four key stages:

High-Fidelity Data Generation: A dataset of 962 simulation cases was generated using COMSOL Multiphysics. The physical model represents a three-segment serial pipeline system with varying diameters (0.106 m, 0.1524 m, 0.2032 m) constructed from AISI 1020 steel. The crude oil was modeled as a non-Newtonian viscoplastic fluid using the Bingham–Papanastasiou regularized formulation to capture yield stress and plastic viscosity effects. The simulations solved coupled continuity, momentum, and energy equations under laminar flow conditions.
Physics-Informed Feature Engineering: To ensure physical interpretability and improve learning efficiency, dimensionless features were engineered from the simulation outputs. These included the Reynolds number ( $Re$ ), Bingham number ( $Bi$ ), and a Thermo-Hydraulic Ratio ( $THR$ ). Crucially, the study introduced a novel Rheological Heat-Loss Index (RHLI), defined as the ratio of viscous dissipation to yield-stress-controlled dissipation. This feature links fluid rheology directly to thermal behavior.
Machine Learning and Ensemble Modeling: A suite of algorithms was trained and cross-validated (5-fold) on the dataset, including regularized linear models (Ridge, Lasso), tree-based ensembles (Random Forest, Extra Trees, XGBoost, LightGBM, Gradient Boosting), Support Vector Regression (SVR), and Neural Networks (MLP). To enhance robustness, Voting and Stacking ensemble strategies were employed, where the Stacking Ensemble served as a meta-learner to combine base model predictions.
Uncertainty Quantification (UQ) and Explainability (XAI): The framework incorporated rigorous UQ using Quantile Regression (estimating 10th–90th percentiles), Conformal Prediction (for distribution-free prediction intervals), and Calibration Analysis. SHapley Additive exPlanations (SHAP) were utilized to interpret model decisions and identify dominant physical drivers.

Key Contributions
The paper identifies five primary contributions:

Dataset Creation: Generation of a comprehensive, physics-consistent thermo-hydraulic dataset for multi-segment Omani crude oil pipelines under varying geometric and rheological conditions.
Multi-Target Surrogate Modeling: Development of AI models capable of simultaneously predicting Heat Loss, Head Loss, and Power Loss, addressing a gap in existing single-target literature.
Novel Feature (RHLI): Introduction of the Rheological Heat-Loss Index as a dimensionless, interpretable feature that characterizes the interplay between plastic viscosity, yield stress, and thermal loss.
Integrated UQ and XAI: Deployment of a framework that not only predicts values but also provides statistically valid confidence intervals and physical explanations for those predictions.
Performance Benchmarking: A comparative evaluation demonstrating that tree-based ensembles and stacking methods significantly outperform linear models and standard neural networks for these specific non-linear thermo-hydraulic tasks.

Results

Predictive Accuracy: The models achieved exceptional performance, particularly for Heat Loss, where tree-based ensembles and the Stacking Ensemble achieved a cross-validated $R^2 \approx 1.0$ . For the more non-linear Head Loss and Power Loss targets, the Stacking Ensemble provided the most stable performance, with $R^2$ values approaching unity.
RHLI Regime Analysis: The analysis of the RHLI revealed three distinct operating regimes: a mixed regime (Low RHLI), a hydraulically dominated regime (Medium RHLI), and a thermally dominated regime (High RHLI). This confirmed the index's utility as an engineering indicator for identifying flow states.
Uncertainty Quantification: The UQ framework demonstrated high reliability. The Mean Calibration Error (MCE) was 0.0211, and the empirical 95% coverage reached 97.4%, indicating that the prediction intervals accurately reflected the true uncertainty.
Explainability: SHAP analysis identified Ambient Temperature as the dominant driver of Heat Loss, a finding consistent with established heat-transfer physics. This validation confirms that the AI models learned physically meaningful relationships rather than statistical artifacts.

Significance and Claims
The paper claims that the developed framework offers a fast, interpretable, and uncertainty-aware surrogate capable of replacing computationally expensive CFD simulations. By reducing prediction time from hours (CFD) to milliseconds (AI), the framework enables:

Real-time Digital Twin Monitoring: The ability to instantly estimate energy losses and identify operating regimes based on sensor data.
Risk-Aware Decision Making: The provision of confidence intervals allows operators to make decisions with a quantified understanding of prediction reliability.
Energy Efficiency: The identification of dominant loss drivers (e.g., ambient temperature) supports targeted operational adjustments to minimize pumping power and thermal dissipation.

The authors conclude that while the current model is trained on simulation data for a specific three-segment configuration, it establishes a robust pathway toward intelligent, energy-efficient pipeline management and real-time monitoring systems for Omani crude oil infrastructure. Future work is noted as necessary to validate these models against field-scale measurements and to adapt them for different rheological formulations.

Technical Summary: Physics-Informed AI Surrogate Modelling of Thermal and Hydraulic Losses in Omani Crude Oil Pipelines

More like this