HARPPP: Autonomous Geometric Design Optimisation of Stirred Tank Reactor Impellers and Baffles

TL;DR

The paper tackles slow, ad hoc geometry design for stirred-tank internals and introduces HARPPP, a simulator-in-the-loop autonomous design framework that couples a compact parametric CAD kernel with power-controlled CFD and CMA-ES optimization. By treating the simulator as the objective, HARPPP explores a 23-dimensional design space and conducts 3,000 autonomous design–simulation cycles, identifying multiple impeller and baffle families that surpass the Rushton baseline in both mean turbulent dissipation $mean(\varepsilon)$ and its spatial variability $CoV(\varepsilon)$ under a fixed power $P=3024$ W, including twisted-plate impellers and curved baffles that define an intensity–uniformity Pareto frontier. The approach generalizes beyond the case study to other equipment and physics models, providing a transparent, auditable platform that retains human-in-the-loop decision-making while avoiding single-point optima. Practically, this enables manufacturable, platform-ready design exploration for industrial internals under constraints such as CIP, GMP, and retrofit feasibility, potentially reducing CAPEX/OPEX while improving process performance.

Abstract

Designing and optimising the geometry of industrial process equipment remains slow and still largely ad hoc: engineers make small tweaks to one standard shape at a time, build prototypes, and hope for gains. We introduce HARPPP, an autonomous design loop that couples a compact, programmable geometry model to power-controlled CFD and evolutionary search. The geometry model is a single mathematical description that reproduces every standard impeller as a special case while spanning an unlimited set of manufacturable shapes. Working with Johnson Matthey on an industrial vessel, HARPPP explored a 23-parameter impeller-baffle space at constant power (3024 W), executing 3,000 simulation cycles in 15 days. The search uncovered multiple design families that outperform a Rushton/4-baffle baseline in both mixing intensity and uniformity, including twisted-plate impellers and pitched/curved baffles (intensity +18 to +78 percent; uniformity CoV -16 to -64 percent). A clear intensity-uniformity Pareto frontier emerged, enabling application-specific choices. Because HARPPP treats the simulator as the objective, it generalises to other equipment wherever credible physics models exist.

Figures (6)

• Figure 1: Architecture of HARPPP: Highly Autonomous Rapid Prototyping for Multi-Phase Processes. HARPPP integrates parametric geometry generation, automated meshing, transient CFD simulation with model-inversion power control, and global optimisation in a closed-loop workflow. Design candidates are created through a programmed CAD kernel (forming a mathematical description of a geometry) and meshing tools before being simulated in OpenFOAM. Power control ensures fair comparison between geometries by maintaining constant energy input per volume. Post-processing evaluates turbulent dissipation and its uniformity, which are combined into scalar optimisation targets. An augmented CMA-ES evolutionary strategy (ACCES) generates new candidate geometries, with the entire cycle executed in parallel on local or cluster resources. Panels on the right illustrate representative impeller geometries and corresponding velocity fields explored during optimisation, showing their design diversity.
• Figure 2: Validation of the CFD framework for stirred tank simulations. (a,b) Representative cross-sectional and top views of the computational mesh used to resolve the baffled stirred tank geometry for the base case. (c,d) Mesh-independence study showing the time evolution of turbulent dissipation rate ($\varepsilon$, solid lines, left axis) and its coefficient of variation (CoV, dashed lines, right axis), together with the corresponding power draw, for coarse ($4.3\times 10^5$ cells), medium ($6.2\times 10^5$ cells), and fine ($8.5\times 10^5$ cells) meshes. (e,f) Timestep-independence study showing $\varepsilon$, CoV, and power draw for timesteps $\Delta t=0.001$-$0.002$ s. Both $\varepsilon$ and power stabilise with negligible variation across mesh and timestep refinements, confirming numerical convergence and validating the framework for subsequent optimisation studies.
• Figure 3: Baseline Rushton turbine performance at constant power input. (a) Turbulent dissipation rate ($\varepsilon$) and (b) velocity magnitude fields in the baffled stirred tank at steady state, showing strong localised dissipation in the impeller discharge stream. (c) Dynamic power-control behaviour: the impeller rotation rate (blue) is continuously adjusted to maintain a target input power of 3024 W (red). (d) Evolution of mean $\varepsilon$ (blue) and its coefficient of variation (red), showing stabilisation within 10 s. These results establish the baseline hydrodynamics against which optimised geometries are compared.
• Figure 4: Impeller optimisation landscape and representative families.Left: Scatter of all evaluated designs in the 23-parameter impeller space, plotted by volume-averaged turbulent dissipation $\mathrm{mean}(\varepsilon)$ ($\mathrm{m}^2\,\mathrm{s}^{-3}$) versus its coefficient of variation $\mathrm{CoV}(\varepsilon)$ (dimensionless). Better performance lies to the right (higher dissipation) and downward (greater uniformity). Later generations are shown as darker points. The dashed curve marks the non-dominated (Pareto) front. The baseline Rushton configuration is shown as a red “+”, and four exemplars (Designs 565, 1066, 1145, 1368; crosses) are highlighted. Right: Corresponding geometries and instantaneous $\varepsilon$ fields (cross-section and top view; colour bar 0--5 $\mathrm{m}^2\,\mathrm{s}^{-3}$), illustrating distinct design families spanning the intensity--uniformity trade-off.
• Figure 5: Baffle optimisation landscape and representative families.Left: Scatter of all evaluated baffle designs, shown in the plane of volume-averaged dissipation $\mathrm{mean}(\varepsilon)$ versus $\mathrm{CoV}(\varepsilon)$. Higher intensity lies to the right; greater uniformity downward. Later generations are shown in darker colours. The red dashed curve denotes the non-dominated set. The baseline configuration is marked with a red “+”; five exemplars (designs 98, 976, 554, 160, 787) are highlighted. Right: Corresponding baffle geometries (empty tank shown for clarity), illustrating the morphological diversity that attains near-front performance.