Identifying fluid flow regimes in rough fractures is essential for subsurface engineering applications involving fractured rocks. However, developing generalizable, accurate, and efficient models for flow-regime identification and nonlinearity classification remains challenging because it requires extensive, diverse datasets spanning broad parameter ranges to ensure applicability across scenarios. To address this, 270 representative synthetic fractures were generated using power spectral density (PSD) methods combined with orthogonal design, resulting in a foundational dataset of 1,890 samples obtained from numerical flow simulations. Different machine learning approaches were then employed to identify flow regimes and partition the degree of nonlinearity according to the non-Darcy effect factor (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$E$</EquationSource> <EquationSource Format="MATHML"><math> <mi>E</mi> </math></EquationSource> </InlineEquation>). To overcome data limitations and enhance the machine learning model performance, an effective physical mechanism-driven data augmentation approach was proposed. This method generates additional data by leveraging the hydraulic gradient, the most critical key factor for nonlinear flow, and its relationship with <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$E$</EquationSource> <EquationSource Format="MATHML"><math> <mi>E</mi> </math></EquationSource> </InlineEquation> in the Forchheimer equation, thus avoiding additional time-consuming simulations. This augmentation strategy enhanced the accuracy of machine learning classifiers and showed strong practicality, excelling in both multi-class classification tasks and simplified key-parameter models. Importantly, the key factors identified via SHAP and MDI are readily obtainable in field applications, ensuring practical feasibility. This study provides a framework for constructing generalizable classification models capable of identifying nonlinear flow thresholds and partitioning nonlinearity degrees across practical parameter ranges, while also introducing an effective framework to enhance machine learning performance through mechanism-driven data augmentation.Highlights<UnorderedList Mark="Bullet"> <ItemContent> Hundreds of synthetic rough fractures are representative for developing generalizable flow-regime identifying and nonlinearity degree partition models. </ItemContent> <ItemContent> Mechanism-driven data augmentation notably enhances machine learning classifiers in multi-class nonlinearity classification tasks. </ItemContent> <ItemContent> The key factors controlling flow regimes in rough fractures are identified and are readily obtainable in practical applications. </ItemContent> <ItemContent> The key-factor multi-class support vector machine models with fewer impact factors developed via data augmentation are precise and highly practicable. </ItemContent> <ItemContent> An effective framework is proposed for constructing generalizable machine learning models enhanced by leveraging physical mechanisms to augment data. </ItemContent> </UnorderedList>

Establishing Generalizable and Practicable Models of Flow Regime Classifying via Mechanism-Enhanced Machine Learning for Various Rough Fractures

期刊论文
发表日期：

2026年4月1日

查看全文

Cixiao Qu,
Mingnan Sun,
Mingyu Wang

摘要

Identifying fluid flow regimes in rough fractures is essential for subsurface engineering applications involving fractured rocks. However, developing generalizable, accurate, and efficient models for flow-regime identification and nonlinearity classification remains challenging because it requires extensive, diverse datasets spanning broad parameter ranges to ensure applicability across scenarios. To address this, 270 representative synthetic fractures were generated using power spectral density (PSD) methods combined with orthogonal design, resulting in a foundational dataset of 1,890 samples obtained from numerical flow simulations. Different machine learning approaches were then employed to identify flow regimes and partition the degree of nonlinearity according to the non-Darcy effect factor ( $E$ $E$ ). To overcome data limitations and enhance the machine learning model performance, an effective physical mechanism-driven data augmentation approach was proposed. This method generates additional data by leveraging the hydraulic gradient, the most critical key factor for nonlinear flow, and its relationship with $E$ $E$ in the Forchheimer equation, thus avoiding additional time-consuming simulations. This augmentation strategy enhanced the accuracy of machine learning classifiers and showed strong practicality, excelling in both multi-class classification tasks and simplified key-parameter models. Importantly, the key factors identified via SHAP and MDI are readily obtainable in field applications, ensuring practical feasibility. This study provides a framework for constructing generalizable classification models capable of identifying nonlinear flow thresholds and partitioning nonlinearity degrees across practical parameter ranges, while also introducing an effective framework to enhance machine learning performance through mechanism-driven data augmentation.

Highlights

Hundreds of synthetic rough fractures are representative for developing generalizable flow-regime identifying and nonlinearity degree partition models.

Mechanism-driven data augmentation notably enhances machine learning classifiers in multi-class nonlinearity classification tasks.

The key factors controlling flow regimes in rough fractures are identified and are readily obtainable in practical applications.

The key-factor multi-class support vector machine models with fewer impact factors developed via data augmentation are precise and highly practicable.

An effective framework is proposed for constructing generalizable machine learning models enhanced by leveraging physical mechanisms to augment data.