Periodic Reporting for period 1 - SCALE (Industry empowerment to Multiphase fluid dynamics simulations using Artificial intelligence and Statistical methods on modern hardware architectures at Scale)
Reporting period: 2023-10-01 to 2025-09-30
SCALE develops optimisation strategies to address real-world industrial applications aligning with the objectives of the European Green Deal and the Horizon Europe mission on Cancer.
The Doctoral Candidates develop data-driven numerical CFD models to predict the behaviour of multiphase flows. This prediction is essential for developing sustainable and innovative technologies with the help of CFD applications.
The industrial applications addressed in SCALE are hydraulic turbomachines, hydrodynamic propulsion, decarbonisation strategies for transportation, pharmaceutical industries, additives, heat transfer and thermal management concepts for electric motors (e-motors).
The research topics are divided into three Work Packages:
WP1: Physics-based and data-driven wall treatment models for non-Newtonian fluids and heat transfer effect
WP2: Physics-informed data-driven surrogate models of complex SGS processes in multi-phase flows
WP3: Consistent data-driven optimization approaches for high-order nonlinear discretisation methods
The Doctoral Candidates develop data-driven numerical CFD models to predict the behaviour of multiphase flows. This prediction is essential for developing sustainable and innovative technologies with the help of CFD applications.
The industrial applications addressed in SCALE are hydraulic turbomachines, hydrodynamic propulsion, decarbonisation strategies for transportation, pharmaceutical industries, additives, heat transfer and thermal management concepts for electric motors (e-motors).
The research topics are divided into three Work Packages:
WP1: Physics-based and data-driven wall treatment models for non-Newtonian fluids and heat transfer effect
WP2: Physics-informed data-driven surrogate models of complex SGS processes in multi-phase flows
WP3: Consistent data-driven optimization approaches for high-order nonlinear discretisation methods
The project advanced data-driven modeling and high-fidelity simulation of turbulent, multiphase, and thermal flows, integrating machine learning across the three work packages.
WP1 focused on wall modeling and turbulent flow simulations using a Deep Reinforcement Learning (DRL) framework coupled with an in-house CFD solver, enabling two-way interaction between simulations and ML-based boundary conditions. Extensive studies of turbulent, viscoelastic, and shear-thinning flows investigated near-wall physics, grid independence, and minimal channel configurations. Post-processing tools generated ML-ready databases. High-fidelity DNS and LES simulations were validated against experiments and reference simulations, producing an open-source LES database for zero- and adverse-pressure-gradient thermal boundary layers. The framework integrates advanced inflow/outflow treatments, pressure-gradient imposition, and thermal coupling, ensuring full reproducibility and ML readiness. Additional studies addressed multiphase compressible flows, including shock-induced cavitation in non-spherical droplets, enhancing post-processing tools and resulting in a journal paper under review. WP1 delivered validated tools, datasets, and ML-integrated workflows advancing wall modeling and multiphase simulations.
WP2 developed ML-derived subgrid-scale (SGS) surrogate models and advanced numerical tools for multiphase and cavitation-driven flows. A multi-fidelity Gaussian Process surrogate predicted maximum wall pressures during bubble-cluster collapse using data across grid resolutions and simplified single-bubble cases, demonstrating strong cross-fidelity correlations. Open-source releases included JAX-Fluids for fully resolved bubble-cluster simulations and the AMReX-based 3phaseSolver for pressure-wave propagation in multi-material bodies. Additional progress included ML-training databases, reproducible agglomerate geometries for medical applications, ultrasound-driven bubble dynamics in soft tissues with real-fluid thermodynamics and fluid–structure interaction, and continuous-adjoint optimization for unsteady VoF multiphase flows with phase change. WP2 delivered validated frameworks, surrogate models, and open-source tools significantly enhancing multiphase and cavitation modeling.
WP3 targeted open-source, fully differentiable solvers and ML-enhanced tools for cavitating flows. JAX-Fluids was extended for DNS of cavitation, incorporating level-set generation from 3D STL and 2D DXF geometries, and weakly compressible and barotropic models, validated on a 3D cavitation tunnel and released in Deliverable D3.1. Complementary work included data-driven modeling for cavitating propellers and complex geometries, development of a boundary graph neural network for 2D airfoil pressure prediction, and pooling strategies for Graph-U-Net architectures applied to 3D ship hulls. Machine-learning solvers based on PINNs and ConvLSTM were explored, with a validated single-phase PINN solver extended toward two-phase cavitating flows and ConvLSTM-based surrogate solvers accelerating iterative CFD schemes via temporal correction learning from coarse-grid data.
Overall, the project delivered validated, reproducible, and open-source tools, high-fidelity datasets, ML-integrated solvers, and surrogate models, establishing a robust, extensible platform for state-of-the-art simulation and modeling of turbulent, multiphase, and cavitating flows. These results enable advanced research, optimization, and industrial applications in thermal systems, propulsion, biomedical therapies, and fluid–structure interactions, providing a scalable framework for combining machine learning with high-fidelity CFD.
WP1 focused on wall modeling and turbulent flow simulations using a Deep Reinforcement Learning (DRL) framework coupled with an in-house CFD solver, enabling two-way interaction between simulations and ML-based boundary conditions. Extensive studies of turbulent, viscoelastic, and shear-thinning flows investigated near-wall physics, grid independence, and minimal channel configurations. Post-processing tools generated ML-ready databases. High-fidelity DNS and LES simulations were validated against experiments and reference simulations, producing an open-source LES database for zero- and adverse-pressure-gradient thermal boundary layers. The framework integrates advanced inflow/outflow treatments, pressure-gradient imposition, and thermal coupling, ensuring full reproducibility and ML readiness. Additional studies addressed multiphase compressible flows, including shock-induced cavitation in non-spherical droplets, enhancing post-processing tools and resulting in a journal paper under review. WP1 delivered validated tools, datasets, and ML-integrated workflows advancing wall modeling and multiphase simulations.
WP2 developed ML-derived subgrid-scale (SGS) surrogate models and advanced numerical tools for multiphase and cavitation-driven flows. A multi-fidelity Gaussian Process surrogate predicted maximum wall pressures during bubble-cluster collapse using data across grid resolutions and simplified single-bubble cases, demonstrating strong cross-fidelity correlations. Open-source releases included JAX-Fluids for fully resolved bubble-cluster simulations and the AMReX-based 3phaseSolver for pressure-wave propagation in multi-material bodies. Additional progress included ML-training databases, reproducible agglomerate geometries for medical applications, ultrasound-driven bubble dynamics in soft tissues with real-fluid thermodynamics and fluid–structure interaction, and continuous-adjoint optimization for unsteady VoF multiphase flows with phase change. WP2 delivered validated frameworks, surrogate models, and open-source tools significantly enhancing multiphase and cavitation modeling.
WP3 targeted open-source, fully differentiable solvers and ML-enhanced tools for cavitating flows. JAX-Fluids was extended for DNS of cavitation, incorporating level-set generation from 3D STL and 2D DXF geometries, and weakly compressible and barotropic models, validated on a 3D cavitation tunnel and released in Deliverable D3.1. Complementary work included data-driven modeling for cavitating propellers and complex geometries, development of a boundary graph neural network for 2D airfoil pressure prediction, and pooling strategies for Graph-U-Net architectures applied to 3D ship hulls. Machine-learning solvers based on PINNs and ConvLSTM were explored, with a validated single-phase PINN solver extended toward two-phase cavitating flows and ConvLSTM-based surrogate solvers accelerating iterative CFD schemes via temporal correction learning from coarse-grid data.
Overall, the project delivered validated, reproducible, and open-source tools, high-fidelity datasets, ML-integrated solvers, and surrogate models, establishing a robust, extensible platform for state-of-the-art simulation and modeling of turbulent, multiphase, and cavitating flows. These results enable advanced research, optimization, and industrial applications in thermal systems, propulsion, biomedical therapies, and fluid–structure interactions, providing a scalable framework for combining machine learning with high-fidelity CFD.
The project delivers substantial advances beyond the current state of the art by tightly integrating machine learning, high-fidelity CFD, and open-source, differentiable simulation frameworks for turbulent, multiphase, and cavitating flows. A key breakthrough is the seamless coupling of ML methods—ranging from Deep Reinforcement Learning, Gaussian Processes, PINNs, and ConvLSTM models—directly within production-level CFD solvers. This enables adaptive wall functions, SGS models, and surrogate solvers that learn complex near-wall, multiphase, and cavitation physics directly from data, overcoming long-standing limitations of empirical or purely analytical closures.
The project moves beyond isolated algorithmic developments by delivering validated, machine-learning-ready simulation ecosystems. These include open-source, reproducible LES/DNS databases for turbulent boundary layers under arbitrary pressure-gradient and thermal conditions; multi-fidelity surrogate models for bubble-cluster collapse that dramatically reduce computational cost while retaining predictive accuracy; and fully differentiable, GPU-native cavitation solvers capable of automatic differentiation and adjoint optimization in multiphase flows with real-fluid thermodynamics and phase change. The continuous-adjoint formulation for unsteady VoF flows with replaceable equations of state represents a first-of-its-kind capability, enabling gradient-based design and uncertainty quantification in regimes previously inaccessible.
The potential impacts are significant across academia and industry. The results enable faster, more reliable design of thermal systems, propulsion devices, medical drug-delivery technologies, and flow-control strategies by replacing expensive trial-and-error simulations with data-driven, optimizable models. Open-source availability and reproducibility lower barriers to adoption, foster community validation, and accelerate innovation. In the longer term, the project establishes a scalable blueprint for integrating ML and CFD, supporting safer biomedical applications, more efficient energy and transport systems, and a new generation of trustworthy data-driven turbulence and cavitation models.
The project moves beyond isolated algorithmic developments by delivering validated, machine-learning-ready simulation ecosystems. These include open-source, reproducible LES/DNS databases for turbulent boundary layers under arbitrary pressure-gradient and thermal conditions; multi-fidelity surrogate models for bubble-cluster collapse that dramatically reduce computational cost while retaining predictive accuracy; and fully differentiable, GPU-native cavitation solvers capable of automatic differentiation and adjoint optimization in multiphase flows with real-fluid thermodynamics and phase change. The continuous-adjoint formulation for unsteady VoF flows with replaceable equations of state represents a first-of-its-kind capability, enabling gradient-based design and uncertainty quantification in regimes previously inaccessible.
The potential impacts are significant across academia and industry. The results enable faster, more reliable design of thermal systems, propulsion devices, medical drug-delivery technologies, and flow-control strategies by replacing expensive trial-and-error simulations with data-driven, optimizable models. Open-source availability and reproducibility lower barriers to adoption, foster community validation, and accelerate innovation. In the longer term, the project establishes a scalable blueprint for integrating ML and CFD, supporting safer biomedical applications, more efficient energy and transport systems, and a new generation of trustworthy data-driven turbulence and cavitation models.