Sonochemical systems are characterised by complex, multi-parameter interactions between acoustic fields, cavitation bubbles, and radical-driven chemical reactions. These interactions are inherently nonlinear and sensitive to operational variables, posing significant challenges for quantitative modelling and system optimisation. This project aims to advance the mechanistic understanding, predictive modelling, and system-level optimisation of sonochemical systems through a structured approach integrating experimental characterisation, dimensionless analysis, and physics-informed machine learning. The research is organised into three interrelated components.

First, the study made use of a standard ultrasound reaction system to experimentally measure sonochemical activity using three complementary indicators: sonoluminescence (SL), sonochemiluminescence (SCL), and potassium iodide (KI) dosimetry. The results revealed systematic differences in sonochemical response characteristics, with both SCL and KI dosimetry as quantitative indicators that reflect distinct reaction mechanisms and spatial sensitivities within the cavitation field. A new image processing method was introduced and validated. The method enhanced spatial resolution and quantitative accuracy, and demonstrated robustness in quantifying SL and SCL signals across a broad range of frequencies and experimental conditions by enabling selective region-based analysis. Second, a set of dimensionless groups was systematically derived from fundamental principles of bubble dynamics and reaction kinetics to capture the underlying physics of sonochemical processes. This marks one of the first attempts in sonochemistry to integrate dimensionless analysis with regression modelling, resulting in compact and physically interpretable models for SCL and KI indicators. The framework improves predictive robustness across varying experimental conditions and provides a generalisable methodology for mechanistic understanding and optimisation of ultrasonic systems. Finally, a machine learning strategy using CatBoost was implemented to improve prediction accuracy and enable model interpretability. Both dimensional and dimensionless input strategies were examined and compared to evaluate their predictive efficacy and physical relevance. SHAP analysis was used to interpret feature contributions and reveal underlying physical decision patterns of the model. This constitutes the first comprehensive application of interpretable machine learning in sonochemistry, demonstrating that physically derived input variables can significantly enhance predictive performance and generalisability. The approach bridges data-driven prediction with mechanistic insight, offering a novel pathway for modelling nonlinear acoustic phenomena.

This thesis introduces a novel, integrative methodology that bridges experimental measurement, physical modelling, and machine learning within a unified framework. It represents one of the first general efforts to incorporate systematically dimensionless physical variables into interpretable machine learning for sonochemistry, enabling predictive precision and mechanistic understanding. The findings contribute new insights into modelling chaotic systems and lay a theoretical and methodological foundation for future research in sonochemical process optimisation.