"We're in the midst of a revolution", says Andreas Adelmann, head of the Laboratory for Scientific Computing and Modelling (LSM) at the Paul Scherrer Institute PSI. By that, the engineer and mathematician means the latest advances in the field of computer technology. "At present we are in a phase of development in which computers no longer need to be programmed by people, but can learn on their own", Adelmann explains. "This mode of operation is similar to that of the brain, which learns in many different ways by processing impressions from outside." The same thing happens in so-called machine learning, except that here the external impressions are replaced by data.
The new approach, however, only complements the traditional approach that Adelmann and his colleagues typically follow. Inspired by theory, the researchers construct a model. This means they design a computer program as a model and use it to carry out simulations.
All models and simulations start with simplification. To obtain a model, scientists describe reality by means of equations. To mimic more complex relationships – a whole system – many such equations are required. These are translated into a mathematical language, the language of computers. The result is a so-called numerical model.
The more complicated the system to be mapped onto a model, the more complex the solution of the numerical model becomes. For this the researchers need high computing power, which is available to them for example with "Piz Daint". That is not only the name of a Swiss mountain peak, but also one of the highest-performance supercomputers in the world, located at the Swiss National Supercomputing Centre (CSCS) in Lugano. But the researchers need not resort to Piz Daint for all of their calculations. "PSI has just invested in upgrading its mainframe computer. At our 'Merlin' we can solve medium-sized problems, for example calculations for SwissFEL, right here in Villigen", Adelmann says.
Saving time and money
Thanks to the improved performance of computers, today researchers can develop models and simulations of systems and processes that cannot be verified in reality: black holes, for example, those enormous, all-consuming mass accumulations in space, which space probes could never approach close enough to carry out the corresponding measurements. There are other events no one ever wants to experience in reality, for example hazardous incidents in complex nuclear power plants. Here too, models and the resulting programs are used for simulation.
With the models they design, the researchers can calculate how an experiment is likely to run, in order to identify potential problems in the test setup. "By simulating the instruments in advance or optimising them in real time, we can make optimal use of expensive beamtime for the actual experiment", Adelmann explains.
Limits are challenges
"We usually achieve a high degree of accuracy with our models and simulations", Adelmann asserts. "Our computational models are like a microscope. We can resolve a great many details when we simulate, for instance, an entire accelerator or a nuclear reactor."
As far as the models and simulations may suffice, even the experts sometimes hit their limits, for instance when chaotic conditions prevail in a system. Then one small change can have a huge, unpredictable impact on the events and their outcome.
One example of this is the dynamics of fluid systems, used to elucidate weather and ocean currents as well as cooling systems. As soon as turbulence arises, things get very complicated. If water flows past an obstacle, eddies form; the smallest factors can influence which side of the obstacle the water then flows past. Such a system can be so sensitive to the slightest changes that multiple runs of the simulation yield completely different results.
"Such systems always highlight the limits of what we can model, or what we just don't understand yet", Adelmann says. Machine learning could be the solution in some cases. "With this new approach, we try to create a computer model of processes in which we use the data of the experiment to learn. The resulting programs are much faster than those programmed by following the conventional path", Adelmann explains. "Here at PSI we are getting acquainted with this new technology and starting to integrate it into our projects."
Text: Paul Scherrer Institute/Christina Bonanati