Supercomputers are extremely versatile and can be utilized in many fields of science and research. In this chapter, we discuss some typical examples of applications of supercomputers in more detail.
The video shows a simulation of gold nanoparticles colliding into a gold surface. The size of the particle is about 20 nm, i.e. a 20 millionth part of a millimeter. Understanding this type of processes is important e.g. for radiation damage on a spacecraft. Macroscopic cratering, for instance by meteorites impacting the ground, can intuitively be approached as a purely macroscopic interaction between two solids, causing stresses and deformation. Conversely, microscopic cratering by atoms can be understood as an atomic process inducing the successive displacement of a limited number of individual atoms. Somewhere between the two extremes, the familiar macroscopic landscape must emerge from microscopic processes. The overlap between phenomena at scales across several orders of magnitude necessitates considerable computing power to simulate, which can only be provided by parallel computing and supercomputers.
When people breathe, they emit aerosols into the air. The spreading of these aerosols can be studied with supercomputers, which in turn helps us in fighting the COVID pandemic. The aerosol particles are vanishingly small and light, so that they can be regarded as points being moved around by air. The motion of the air itself is a very complex thing to simulate. Even in seemingly gentle conditions, such as blowing away the vapour over a cup of coffee or the smoke of a candle, the motion of the air is visibly disordered, or in technical terms turbulent. Turbulence is extremely challenging to simulate, because a turbulent flow is a superposition of vortices at all scales. Mathematician Lewis Fry Richardson put it in more poetic terms:
Big whirls have little whirls Which feed on their velocity, And little whirls have lesser whirls And so on to viscosity.
Simulating the transport of aerosol into a room needs to scale from the geometry of the room itself, all the way down to the minuscule vortices that barely survive the dissipative action of viscosity.
In order to understand the climate effects of carbon emissions, simulations over long time scales are needed. The video below shows a millenium-scale simulation concerning the surface air temperature in the Arctic and the Greenland glacier ice thickness, answering questions such as when the Arctic ice is gone during the summer and winter. In this case, not only does the system involve very disparate scales (in time as well as in space), it also implies the coupling of systems governed by very different physical models, each of which demands substantial computing power in its own right. Here a model for the evolution of the ice cover and one for the Earth's atmosphere and oceans are simulated in lockstep, providing each-other with time-dependent parameters at the interface, such as snow precipitation from the atmosphere to the glacier, and fresh water influx into the ocean due to the melting of the glacier.
Neural networks and deep learning have enabled important advances in the processing of language. By uncovering correlations between words or clusters of words, neural networks are able to learn underlying elements of language such as lexical context. Training such neural networks requires an enormous amount of text, which is often obtained from sources such as Wikipedia, news articles, or forum discussions among others. The processing of such a large corpus text is enabled by supercomputers, in particular the more recent machines comprised of many GPU nodes.
An artificial intelligence system can identify biopsies containing cancer nearly without error. Researchers digitally scanned more than 8,000 prostate biopsies to train and test the artificial intelligence. An artificial intelligence system consisting of ten deep neural networks was trained to distinguish between benign and cancerous prostate biopsies.
Approximately six million images extracted from digitally scanned biopsies were used to train the artificial intelligence. There were approximately 30 terabytes of image data, and the final training data consisted of more than 2 trillion pixels. With a normal single CPU, it would have taken at least several months if not years to train the model, but a supercomputer with GPUs dropped the total computing time to 2-3 days.
The goal of the artificial intellisgence system is not to replace human experts, but rather to provide pathologists with a tool that can, on one hand, improve work efficiency, while also promoting patient safety by acting as a safety mechanism.
Due to extreme versatility of supercomputers, exhaustive listing of all important applications is not feasible. Below you can, however, find information on some additional use cases.
- Topological superconductor – new building block for qubits
- Understanding biological energy conversion via molecular simulations
- Operation of nanocatalyst at the atomic level
- Structural biology guiding development of new SARS-Cov2 therapies
- Digging deeper in dynamics of ice with glacier simulations
- Visualization speaks more than a thousand numerals