EUVIMEDEUVIMED entdecken
Crossref

From composition to deposition: A machine learning framework for wax deposition prediction in petroleum fluids using compositional properties

Mohammadali Ahmadi

Physics of Fluids · 01.07.2025 · 37 (7)

DOI: 10.1063/5.0276315 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Physics of Fluids, online erschienen am 30.07.2025. Verfasst von Mohammadali Ahmadi. Die Arbeit wurde bisher 13-mal zitiert (Crossref) und nennt 43 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Wax deposition in petroleum pipelines presents a significant flow assurance challenge, requiring accurate predictive models to optimize mitigation strategies. This study systematically evaluates three machine learning approaches, random forest, extreme gradient boosting (XGBoost), and physics-informed neural networks (PINNs), for wax deposition prediction using comprehensive fluid compositional and thermodynamic data. A curated dataset of 88 experimentally measured samples was used, including detailed hydrocarbon composition (from C1 to C30+), temperature, pressure, and observed wax deposition percentages. Key thermodynamic descriptors such as the weighted average carbon number and long-chain hydrocarbon ratio were engineered to capture relevant physical behavior. Through comparative performance analysis, uncertainty quantification, and interpretability diagnostics, we demonstrate that XGBoost achieves superior predictive accuracy [R2 = 0.813–0.957; mean absolute error = 0.095–0.258] while maintaining robust generalization capabilities. The model's gradient-boosting architecture effectively captures nonlinear relationships between critical factors, including temperature and hydrocarbon fractions (C1–C3, C8–C15, and C16–C22), and wax deposition behavior. Partial dependence and SHAP (SHapley Additive exPlanations) analyses reveal physically interpretable patterns: temperature dominates as an inverse predictor, while mid-range hydrocarbons (C8–C15) exhibit inhibitory effects. Beyond accuracy, this study addresses critical gaps in the existing literature by incorporating uncertainty quantification via bootstrap resampling and model explainability using SHAP. Unlike previous efforts that relied on point prediction or lacked physical integration, this work demonstrates how hybrid data-physics modeling can enhance robustness and decision confidence. Although PINNs incorporate valuable thermodynamic constraints, they underperform data-driven models due to systematic biases in balancing physical priors with empirical learning. Bootstrap uncertainty analysis confirms XGBoost's operational reliability, with well-calibrated prediction intervals across diverse scenarios. These findings establish XGBoost as the preferred method for industrial wax prediction, while highlighting opportunities for hybrid physics-informed machine learning approaches. The study provides both practical guidelines for pipeline management and a framework for integrating domain knowledge into predictive modeling of wax deposition.

Zitieren

APA

Ahmadi, M. (2025). From composition to deposition: A machine learning framework for wax deposition prediction in petroleum fluids using compositional properties. Physics of Fluids, 37(7). https://doi.org/10.1063/5.0276315

Vancouver

Ahmadi M. From composition to deposition: A machine learning framework for wax deposition prediction in petroleum fluids using compositional properties. Physics of Fluids. 2025;37(7). doi:10.1063/5.0276315

BibTeX
@article{Ahmadi2025,
  author = {Ahmadi, Mohammadali},
  title = {From composition to deposition: A machine learning framework for wax deposition prediction in petroleum fluids using compositional properties},
  journal = {Physics of Fluids},
  year = {2025},
  volume = {37},
  number = {7},
  doi = {10.1063/5.0276315}
}
Kostenlos starten →