Multiscale Simulation and Machine Learning Optimization of Cuttings Transport in Riserless Pipelines Under Bubble Coalescence-Breakup Driven Flow.

Abstract

In deepwater riserless drilling, inefficient cuttings transport remains a critical challenge, leading to incidents such as stuck pipes and lost circulation. While existing studies have primarily focused on liquid-solid two-phase flow, they largely overlook the nonlinear coupling effects of bubble dynamics (coalescence and breakup) in gas-liquid-solid three-phase systems, resulting in poor predictive accuracy under high-pressure, high-viscosity conditions. This study introduces a systematically integrated framework that presents a novel integration of visualization experiments, multiscale numerical simulation (CFD-DEM coupled with the population balance model, PBM), and machine learning. This methodology provides an in-depth understanding of the bubble coalescence and breakup mechanisms in cuttings bed formation and transport efficiency under riserless drilling conditions. The coupled CFD-DEM-PBM model provides high-fidelity quantification of bubble dynamics and its cross-scale interaction with cuttings. Subsequently, high-resolution, experimentally validated data are utilized to train both backpropagation (BP) and radial basis function (RBF) neural networks to achieve physics-informed, intelligent prediction of cuttings concentration and transport rate. The results demonstrate that bubble coalescence reorganizes local flow fields to suppress sedimentation, whereas bubble breakup induces microscale turbulence that promotes particle resuspension; their synergy critically governs the overall transport efficiency. Notably, the RBF neural network achieved superior predictive accuracy (<i>R</i> ² = 0.91182), significantly outperforming both the BP network and traditional empirical models. This work not only offers a reliable basis for real-time parameter optimization and decision-making but also establishes a new paradigm for understanding complex multiphase transport phenomena through the seamless integration of high-fidelity simulation, systematic experimentation, and explainable machine learning.