Graph Database Integration and Machine Learning for Operational and Hydrodynamic Analysis of a Liquid Mixing System

ABSTRACT This study evaluates the hydrodynamic behavior of the MF‐TM‐200 liquid mixing system through an integrated approach combining dimensionless analysis, graph‐based representation, and machine learning. More than 1000 experimental configurations corresponding to eight impeller designs were analyzed under different operating conditions, using the Reynolds number (NRe), Froude number (NFr), and dimensionless liquid‐height ratio (Φ) as hydrodynamic descriptors. The experimental data were structured in Neo4j to represent associations among configurations and explore patterns within the experimental space. The graphs revealed differences among the impellers and identified relevant coincidences between Impellers 3 and 8 in NRe and NFr, while Φ provided differentiating conditions related to the relative free‐surface response. The three descriptors were subsequently used to classify impeller type using five machine learning algorithms: gradient boosting machine (GBM), random forest (RF), XGBoost, extra trees (ET), and LightGBM. All models achieved balanced accuracy values above 0.90 and ROC‐AUC values above 0.93. LightGBM showed the best overall performance, with an accuracy of 0.8824, balanced accuracy of 0.9317, and ROC‐AUC of 0.9812. Overall, the integration of graph‐based analysis and machine learning provides a complementary framework for identifying hydrodynamic patterns and discriminating among impeller configurations within the investigated laboratory‐scale mixing system.