Session 7 - Theme: Mixing & Processes (1)

Despite the fact that computational modelling in multiphase flow mixing has aroused interest over the past 40 years or so, its role and usefulness still remain unclear. This presentation will attempt to bring elements of answer to this question by reviewing some of the most significant techniques and methods that have been developed to simulate mixing flow in various applications. Through many examples of industrial relevance, we will discuss issues such as problem setting, model development, model verification and validation, determination of physical properties, hardware resources, and convergence of results. Upon doing so, we will evidence the progress made in computer modelling but also bring up aspects that need to be further improved, such as the availability of robust and efficient techniques that can cope with multiphase and multi-physics mixing flows. This will lay the table for the recent interest in the so-called digital twins, which aim at bringing together in a synergistic manner correlations from the literature, experimental data, AI and computer models. The state-of-the-art concerning the digital twin paradigm within the scope of mixing applications will be discussed as well as its current limitations and challenges to be addressed in the future.

This work introduces an Artificial Intelligence–assisted Radioactive Particle Tracking (RPT-AI) framework for rapid and quantitative characterization of multiphase hydrodynamics in mechanically agitated reactors. The approach combines robotized calibration with deep neural-network reconstruction to directly map detector radiation signals to three-dimensional tracer positions at sub-millimetric accuracy, eliminating the need for model-based inversion or Monte Carlo simulation. Millions of detector–position pairs are collected automatically, allowing the trained network to generalize across geometries and operating conditions without empirical tuning. The resulting flow fields provide velocity distributions and derived metrics including pumping rates, mixing constants, and near-wall velocities within a single-day experimental cycle, at a fraction of the time and cost of CFD or legacy RPT workflows. The framework was applied to stirred tanks while systematically varying impeller type, spacing, and bottom geometry. Analysis at constant power input revealed that multiple optima exist depending on the process objective. Strongest bulk recirculation occurs for dual-Rushton systems with larger inter-impeller spacing or for propellers operating over ellipsoidal bottoms, whereas fastest homogenization is achieved with closely spaced Rushton impellers that reinforce axial exchange. Wall-velocity measurements showed that heat-transfer performance correlates more strongly with near-wall hydrodynamics than with conventional tip-speed surrogates. These results highlight the limits of classical correlations and CFD closures, which reproduce qualitative trends but often underpredict geometry-dependent effects and multi-impeller interactions by 20–30%. By directly measuring local flow fields rather than inferring behavior from surrogate parameters, RPT-AI provides an experiment-based foundation for hydrodynamic scaling and design optimization. The method enables accelerated design–build–test cycles wherein CAD modifications and additive-manufactured prototypes can be evaluated iteratively in real reactors. Beyond characterization, the large high-fidelity datasets produced by RPT-AI create a pathway toward AI-driven hydrodynamic models trained on experimental truth data. These “large process models” may bridge CFD, digital twins, and operations through hybrid, data-validated reactor simulation. RPT-AI therefore establishes a scalable and geometry-agnostic platform for predictive hydrodynamics, supporting faster and more reliable reactor design and scale-up.

This work proposes a new methodology for predicting flow behavior in chemical reactors, applied to the specific case of Confined Imping Jets (CIJs), aiming to significantly reduce computational time and memory requirements. The proposed approach consists of two main steps: (i) dimensionality reduction of CFD data using Proper Orthogonal Decomposition (POD) to extract the dominant flow structures, and (ii) the development of Deep Neural Networks (DNNs) to predict POD time-independent functions and associated time-varying coefficients for new operating conditions. CIJs consist of a confined cylindrical chamber where two opposed streams are injected. For inlet Reynolds numbers above 150, a self-sustainable chaotic flow regime is established. In this work, the Reynolds number was selected as the governing operating parameter and varied between 300 and 600. 2D CFD simulations were performed, and POD was applied to the resulting velocity fields. The POD analysis showed that the first mode represents a steady state flow, while the subsequent modes correspond to large-scale vortical structures. Two independent DNNs were trained: one to predict the spatial modes (time-independent functions) and another to predict the temporal coefficients. The networks were trained, and their performance was evaluated for an unseen condition (Re = 450). The DNN predictions showed very good agreement with the POD data obtained from CFD, enabling an accurate reconstruction of the velocity field. This study represents the first application of DNNs to predict POD data for new operating conditions in CIJs and demonstrates strong potential to accelerate flow-field predictions in chemical reactors.