GPU-Accelerated Simulations Boost Reactive Flow Modelling with Adaptive Meshing.

A new computational framework utilising adaptive mesh refinement and GPU acceleration delivers substantial performance gains for simulating complex reactive systems. Benchmarks confirm fidelity and demonstrate up to 6.49x acceleration for problems like hydrogen detonation, alongside improved computational efficiency across multiple GPU nodes.

The accurate modelling of compressible reactive flows – those involving combustion, detonation, or other chemical processes coupled with fluid dynamics – demands substantial computational resources. Researchers continually seek methods to enhance simulation efficiency without compromising accuracy, particularly for problems exhibiting complex physics and varying timescales. Yuqi Wang, Yadong Zeng, Ralf Deiterding, and Jianhan Liang detail a new heterogeneous adaptive mesh refinement (AMR) framework designed to accelerate these simulations. Their work, entitled ‘An efficient GPU-accelerated adaptive mesh refinement framework for high-fidelity compressible reactive flows modeling’, presents a highly parallelised codebase leveraging Graphics Processing Units (GPUs) to achieve significant speedups in both non-reactive and reactive flow simulations, demonstrated through benchmarks including a three-dimensional reactive shock-bubble interaction problem.

Adaptive Mesh Refinement Framework Enhances Simulations of Reactive Flows

A new computational framework demonstrably improves the efficiency of simulating moderately stiff reactive flow problems. The system integrates adaptive mesh refinement (AMR) – a technique that concentrates computational resources on areas of high activity – with a subcycling-in-time algorithm and a specialised refluxing algorithm, all within a highly parallelised codebase.

AMR dynamically adjusts the mesh resolution during a simulation, using finer meshes where greater detail is required, and coarser meshes elsewhere. This reduces computational cost without sacrificing accuracy. The subcycling-in-time algorithm allows different regions of the simulation domain to advance in time using different time step sizes, optimising efficiency where rapid changes occur. The refluxing algorithm manages the exchange of information between these differing time levels, ensuring stability and accuracy.

The framework prioritises memory efficiency, crucial for GPU-based computations. It achieves this through a low-storage variant of explicit chemical integrators – numerical methods used to solve chemical kinetics equations – reducing the demand on GPU registers and accelerating calculations.

Comparative tests reveal substantial performance gains. The framework achieved speedups of six and three times respectively when compared to implicit and standard explicit methods, while maintaining comparable accuracy. Simulations of hydrogen detonation propagation demonstrated an overall 6.49 times acceleration ratio when utilising a V100 GPU compared to an Intel i9 CPU, within the same codebase.

Scalability tests, conducted using the AMReX infrastructure, confirm the framework’s ability to maintain computational efficiency as computational resources increase. This is vital for leveraging the full potential of modern high-performance computing architectures.

To demonstrate the framework’s capabilities, researchers performed a large-scale direct numerical simulation (DNS) of a three-dimensional reactive shock-bubble interaction. DNS resolves all turbulent scales directly, providing highly accurate but computationally expensive results. This simulation achieved significant cost savings while maintaining accuracy comparable to a prior uniform DNS study conducted on CPUs.

Future development will focus on extending the framework to accommodate more complex chemical mechanisms and incorporating advanced turbulence models. Researchers also plan to investigate the integration of machine learning techniques for both adaptive mesh refinement and chemical kinetic rate prediction, with the aim of further optimising performance and enhancing predictive capabilities. Broadening the framework’s applicability to a wider range of industrial and scientific problems remains a key priority.