Machine learning modelling of sonochemical systems using physically-derived dimensionless groups.

Area of Science:

• Physical Chemistry
• Chemical Engineering
• Data Science

Background:

• Sonochemistry presents complex, nonlinear interactions challenging traditional modeling.
• Current models rely on dimensional variables, limiting their generalizability and interpretability.
• Extrapolation across different sonochemical systems is difficult with existing approaches.

Purpose of the Study:

• To develop a machine learning strategy integrating dimensionless variables (Π-terms) for improved sonochemistry modeling.
• To overcome limitations of conventional models in generalizability and interpretability.
• To provide mechanistic insights into nonlinear sonochemical behaviors.

Main Methods:

• Implemented a machine learning framework using a categorical boosting (CatBoost) algorithm.
• Integrated physically derived dimensionless variables (Π-terms) as input features.
• Evaluated seven supervised learning algorithms, selecting tree-based models for superior performance.
• Utilized SHAP analysis for feature attribution and mechanistic interpretation.

Main Results:

• The machine learning framework achieved high predictive accuracy (R² = 0.87–0.95) on test sets.
• Models using dimensionless inputs generalized to external datasets without corrections, unlike dimensional models.
• Dimensionless models showed superior generalizability and task-to-task consistency, reducing plateau effects.
• SHAP analysis identified cavitation thermal buffering and energy input scaling as key factors (>50% importance).

Conclusions:

• Machine learning with dimensionless variables offers a robust and generalizable approach to sonochemistry modeling.
• This strategy overcomes limitations of dimensional models, enabling accurate predictions across diverse systems.
• The findings provide valuable mechanistic insights into complex nonlinear sonochemical processes.