Mixed precision efficiency improvements for multi-physics flow simulations

In recent years, accelerated architectures (e.g., graphics processing units) have become increasingly prevalent in contemporary high performance computing architectures. The developments in these architectures have been driven by machine learning/artificial intelligence applications where lower precision computing is frequently used. To properly leverage these advanced architectures, scientific computing algorithms must be able to utilize lower precision computing without sacrificing the needed levels of accuracy. This project seeks to use these lower precision calculations specifically applied to high fidelity, multi-physics flow simulations (e.g., noise, turbulent combustion, convective heat transfer). Progress to date has shown that targeted use of lower precision computing with algorithmic modifications can replicate acceptable levels of accuracy for problems involving turbine blade heat transfer, high speed jet noise, and other problems where subtle flow phenomena (e.g., aeroacoustics, laminar-turbulent transition, boundary layer separation) need to be predicted. These improvements have had significant impact on throughput as well enabling computations on lower cost graphics processing units. Additional work is proposed to make similar algorithmic improvements for a wider class of multi- physics flows including particle-laden turbulent flows and liquid fueled combustion processes. Proposed work will also continue the validation of these approaches against available experimental data. Successful completion of this work scope would establish theoretical and empirical justification on the viability of reduced precision operations for high fidelity flow simulations. The deployment of these mixed precision algorithms will significantly reduce the cost of high fidelity flow simulation allowing more commercial use of these tools in the engineering design process.