Using one of the most intense uses of the HiPerGator supercomputer to date by University of Florida engineers, a previously impossibly difficult simulation with home fire safety, heating and cooling applications has been successfully completed.
Because of the complexity of air movements, it has never been possible to simulate thermal wall clouds in such detail. However, the research team led by UF engineering professor Sivaramakrishnan Balachandar was able to track the turbulent air vortices that twist and turn at the submillimetre level, thanks to the dedicated use of 90% of HiPerGator's AI cluster for several days.
According to Balachandar, “We used almost the entire HiPerGator AI cluster to solve a problem that was never resolved in our community at this level of granularity.” “One of the most important problems in science and engineering is turbulent flow. Everywhere we go, whether on planes, on storm tracks or in volcanic clouds, turbulence has an effect on us.”
Thermal wall clouds form when warm, buoyant air rises along a vertical surface. This process occurs during house fires and can spread fires quickly if left unchecked. But less harmful thermal wall clouds occur daily as warm or cooled air rises or sinks along walls in interior rooms. Mudslides and sediment-filled flows – swirling flows – are both explained by very similar processes.
Experiments have been used by many scientists to study thermal fumes, but they are expensive to build and have a limited number of sensors that can be mounted on a wall. These sensors affect the measurements taken, further clouding the data.
Many of the problems encountered in real-world research are solved by computer models of thermal wall fumes, but simulations that can be run on a standard computer are fuzzy and low-resolution. Balachandar's team needed the resources of a powerful supercomputer to achieve millimeter scale.
Airflow patterns in a real dwelling were replicated by the researchers in their simulations. They started the warm air rising virtually from the bottom of a wall along the skirting and observed its change over time. Similar to how thermal clouds can grow in a real house, the model house featured vertical walls and alternating pitched roof lines.
According to Balachandar, these simulations, along with real-world experiments and ideas, are an important component of scientific progress.
“We use computers to tackle Mother Nature, and computer simulation gives us unprecedented access to all the intricacies inside. Using our simulation, we can get inside the wall hair and see every nook and cranny,” explains Balachandar.
The researchers tracked nearly 100 billion elements, including velocity, pressure, and temperature, over a quarter-million instantaneous time period. This task required 140 out of 125 nodes of the HiPerGator AI cluster. Each node has eight GPUs and 128 CPUs, each performing various types of computation. Balachandar's group improved the performance of their simulations by optimizing their code to run on NVIDIA GPUs running AI cluster nodes.
Such comprehensive simulations are also reflected in real-world applications. For example, to construct and understand home heating systems or fire standards, engineers use fairly simple models that may contain erroneous assumptions. These designs can be made better by strengthening these models.
“We can now test existing models and identify where they fall short. We aim to use artificial intelligence to analyze our terabytes of data to help us build better models that others can use,” said Balachandar.