"The art of modelling lies in separating the wheat from the chaff"

Simulate the behaviour of people in crowds: In mathematics and numerics there has been a human
touch lately. Prof. Hans-Joachim Bungartz explains the reasons. Photo: A. Bracken/Unsplash

Happy birthday! Hans-Joachim Bungartz, mathematician and computer scientist, chair of scientific computing at the Department of Computer Science of the TUM School of Computation, Information and Technology at the Technical University of Munich (TUM) and one of the directors of the Leibniz Computer Centre (LRZ), has just turned 60. And it was an excellent occasion to talk to the coding and equations specialist about simulations and developments in numerics. "I like the abstract or generic nature of being able to solve many questions with a numerical method,” he says. I've always been inspired by the fact that as a numerical scientist you are also an enabler". Making things possible - Bungartz has a hands-on, pragmatic, sympathetic and approachable appearance. He wears outdoor sandals instead of leather shoes, and a checked shirt with rolled up sleeves instead of a suit. Bungartz is fast-talking, uses few technical terms and backs up many of his statements with practical examples. Artificial intelligence and data science are now attracting much more attention to his subject: "Numerics used to be a special tool that only a few people needed. Now almost everyone uses numerics, which makes the subject more attractive. Such developments give you the feeling of being in the right place at the right time," he says. In addition, "simulation has become indispensable in industry and science. But it's not just about developing fancy code. Most of the work consists of adapting existing codes so that they can be calculated on a PC or supercomputer in a short time". Recently, a human element has also been added to simulation - namely when it comes to finding formulasfor human behaviour.

Prof. Dr. Hans-Joachim Bungartz: Mathematics, Computer Scientist, LRZ Director

How did you get into computer science and mathematics? Prof. Dr. Hans-Joachim Bungartz: I didn't really know what to do before I finished school - a luxury problem when you have very good grades. So when I was in grade 13, I asked my teachers what the most difficult subject was. They all said maths, but one said it was a bit old-fashioned and that I should try computer science, there were lots of new things happening there. And so it happened.

You worked on algebraic multigrid methods, on thin grids for your PhD, and on the three-dimensional Poisson equation. What do you need that for? Bungartz: My doctoral thesis was one of the early works on thin grids. Multigrid methods are a numerical classic, a way to solve large systems of equations efficiently. When I was studying them, they were finally properly understood and widely used. Thin grids were once intended for similar numerical problems and are now used 80 to 90 per cent in data analysis. In principle, it's madness - a numerical tool anticipates high dimensionality and it is now used for data analysis. In classical simulation you work with continuum, mechanics, physics, three dimensions plus time. Today, we calculate with many dimensions and use data and AI to solve parameter-dependent problems. Many parameters make the high-dimensional exciting. This is quite a typical thing in science: something like thin grids is developed with a clear numerical goal, but it takes 20 years to succeed in a completely different area.

You earned your doctorate with a scholarship from Siemens and were an assistant at the Institute of Computer Science at the TUM. Since then, scientific computing and simulation have been part of your work. What do you find interesting about it? Bungartz: Climate researchers use simulations to answer climate questions, the materials scientist uses them to develop new materials. I'm not really interested in that. I like the abstract or generic, the fact that I can solve many questions with a numerical methodology. It has always inspired me that as a numerical scientist you are also an enabler. Some of my colleagues are frustrated that they are hardly ever listed inpublications about great achievements. Professor Roland Bulirsch, one of the greats in numerics, who sadly passed away last year, said there is numerics in everything and that technology only works on the basis of numbers. But when a robot can do something for the first time, then robotics is celebrated, not the provider of the fundamentals. I don't really care, I like the fact that numerics intersects with many other scientific communities. I have always found the monoculture of the subject narrow.

So you prefer to develop algorithms for the models? Bungartz: Exactly, the models typically come from the domain sciences. Mathematicians have been very helpful, but they are different from numericians. People like us take the model and try to find an efficient solution for use on the computer.

Simulation also takes important questions away from public

You are a co-author of the seminal book "Modelling and Simulation". For which research disciplines do you translate models into codes? Bungartz: By model I mean the mathematical-physical description, the equations that describe a relationship in physics, chemistry or engineering. In scientific computing, we translate models into efficient, easily parallelisable algorithms for computers and supercomputers. And this can keep us busy for years. For example, in plasma physics, when the processes in a plasma reactor are simulated. We have been working with Professor Frank Jenko's group at the Max Planck Institute for Plasma Physics for more than 10 years. The group felt that if they could go one step further with the algorithm, they could do more with their models. So first we changed the solver, then we improved the discretisation and parallelism. Because the standard model is five- or six-dimensional, we rely on thin grids. The discretisation creates points in each direction, but if you want a reasonable resolution you need 10 or better 1,000 or even 10,000 points in each direction, so you have to calculate in dimensions of 10 to the power of 18 or 10 to the power of 24. This is still too much for a supercomputer. With the thin grids, however, the problem can be broken down to the low-dimensional level. But we are far from finished, we need even more efficient algorithms and even better codes.

Where do simulations really make sense? Bungartz: Everywhere, of course. No, seriously, I think, it's a question of economics, a simulation is often more efficient than an experiment, and today it's also possible to make predictions. In the medical field, for example, a lot can be achieved by combining simulation and experimentation. For me - and this may sound tricky - nuclear weapons testing is a good example of the benefits and risks of simulation: It was not the most moral governments that were the first to stop nuclear testing, but those that were the first to simulate on the computer. First the US, then the UK and France, then Russia and China. This also raises an ethical issue: is the world a better place because nuclear weapons are tested on computers? At first glance, yes. The Mururoa Atoll in New Caledonia is no longer disturbed. There are no more television pictures of mushroom clouds anywhere in the world. But of course, weapons technology is still being developed. And this is certainly where the dangers of simulation begin: It takes important questions away from the public. I can't hide important experiments, but when they are simulated, a lot remains hidden.

Music relaxes: Bungartz plays at the Akademischer OrchesterVerband

Where do you get your inspiration? Bungartz: I like being outdoors or playing music. That can be inspiring, but actually I tend to switch off and let my mind wander. Rarely do I experience situations where I think about a scientific problem while running and then, at 4.7 kilometres, something clicks. No - I find inspiration in discourse and conversations with PhD students or colleagues, or when researchers present me with a problem. Then I try to build a bridge to projects and problems that I have already solved with my team.

What developments in recent years have really advanced your work? Bungartz: As computers have evolved, we have been able to solve more and more problems. On the other hand, we are challenged when new computer architectures emerge, like now for the exascale class. Then numerical algorithms have to be adapted to new processors and accelerators. What has also pushed us forward recently is AI and data science. This has brought scientific computing more into the spotlight. Numerics used to be a special tool that only a few people needed. Now a lot of  students use numerics, which makes the subject more attractive. Developments like this give you the feeling of being in the right place at the right time.

Since 2008 you are a member of the borard of directors of the LRZ. Do-you use its supercomputers? Bungartz: We do indeed try out the computers and the supercomputer from time to time. Although we are not the best customers of the big data centres with the fattest jobs, my university department is involved in a number of research consortia in which we test the suitability of codes, evaluate the scaling of programes or calculate algorithms on hundreds of thousands of cores. We often work on the high-performance systems when they are shut down for maintenance. This creates time windows in which we can try out the whole machine.

A certain level of abstraction and expertise

Models describe realities, functionalities, phenomena: What skills do you need to develop good models? Bungartz: For models you need a certain level of abstraction and you should know what the right equations are for each modelling step. You should be well-versed in your subject. In a climate simulation, for example, the question is which physical or chemical effects, i.e. which parameters, should be included. The art of modelling is to separate the wheat from the chaff. Engineering and technology want to use simulation to build better systems. Where is the potential for optimising an engine? The third major task of simulation, which is enhanced by AI, is forecasting: Whether we are talking about natural disasters, the climate or the financial markets, the aim is no longer just to understand why a crisis has occurred, but also to develop instruments to prevent it. This is the ultimate discipline. Seismologists will probably not be able to use simulations to predict exactly where and when earthquakes will occur, but they can calculate probabilities and take appropriate protective measures.

To which students would you recommend model arithmetic and numeracy? Bungartz: At our department, computer scientists and mathematicians mainly do bachelor's or master's theses. Some have a background in physics or engineering. If you work purely theoretically or experimentally, if you want to discover groundbreaking particles, numerics is definitely not the right thing for you. Numericians want to take formulas and codes into applications. To do this, you should be able to immerse yourself in mathematics and computer science and be able to deal with computer architectures. A broad interest in many scientific disciplines and a willingness to take unconventional paths are useful. The prospects are good: simulation is now an integral part of industry and science. But it's not just about developing fancy code. Most of the work involves adapting existing codes so that they can be run on a PC or supercomputer in a short time.

Have you had any role models in your career? Bungartz: My career wasn't planned, it just happened. Perhaps this is not unusual in science. However, I was inspired by the career of my supervisor, Christoph Zenger. He studied physics, did his doctorate in mathematics and then ended up at the department of computer science. He was and still is interested in finding concrete solutions using neat algorithms that require a little more brain power.

You mentioned it earlier - Covid-19, climate change: Why are we so often confronted with model calculations and simulations at the moment? Bungartz: Simulation experienced its first boom when computer technology flourished and mainframe computers became established. That was roughly in the 1980s, and it really took off after 1990. At that time, the automotive industry was hit hard and was looking for ways to reduce development costs. The wind tunnel still exists, of course, but most components are now developed using simulation. When the Mercedes A-Class crashed in 1997, simulations were blamed. Actually, the criticism should have been that they didn't simulate properly, didn't do enough or misinterpreted the simulation results. This first simulation boom arrived in science and business, but not yet in society. The second boom followed a few years ago with the advent of AI and data science. The hope now is that we will be able to go even deeper into forecasting and make reliable predictions - even for questions that are of interest to everyone. Predictive science is a goal driving simulation today. The pandemic has also given modelling a massive boost. There hasn't been a night when someone hasn't been on television talking about what their own models are showing. Simulations are becoming more and more important, and there are also questions that cannot be answered experimentally, for example in astrophysics.

Is the visualisation of simulations something that adds value? Bungartz: If astrophysics makes it onto the 8 o'clock news, it's not necessarily because of great discoveries, but because there are spectacular images. In scientific research, you have to be able to visualise certain things. To evaluate simulations, we used to print out piles of paper, now it's huge files and mostly numbers, but what do they actually mean? Visualisation provides an intense and direct experience of the results and is a key tool for analysing data.

Are there any pitfalls in modelling one can fall into? What are the biggest errors in algorithms? Bungartz: There are many pitfalls. You can - unintentionally or due to the effort required - leave out the most important parameters in a model. People also like to believe what they see. Once three or four examples have been calculated and the desired results have been achieved, it is easy to become too enthusiastic and not test extensively enough. If a code is developed especially for a computer in a data centre, or if you want to get a result quickly, you are not thinking hard enough. Then the algorithm has to be optimised again after a short time.

How are machine learning and AI changing simulation? Bungartz: Classic simulation starts with the model, a deductive process: You take equations, a hypothesis, and derive data from them. Data science now allows you to go the other way: without knowing any models, I can analyse data and then derive a model from it. Analysing messages on social media often works better than physical models. In Japan, for example, Twitter data was found to be a great way to predict earthquakes. Hours before an earthquake, the birds stop singing. People are talking about it on Twitter. In retrospect, this is a very good predictor - an earthquake can be timed and even localised quite accurately. Physical models, on the other hand, show the probability that something will happen, but when, where and how strong it will be is often left open. There are other examples of deriving a mathematical model from observational data. You could imagine putting together all the little films I get of rivers, streams or lakes and having an AI derive a differential equation for currents from them. Of course, it doesn't work that way, but thanks to AI we now have inductive access to models. In this respect, simulation and AI are a bit like ping-pong. One is the way to get from models to data, and the other is the way to get from data to models. And in this interplay, both are going to accelerate the process of gaining knowledge . This is exciting for science, but it requires a good understanding of models and their fundamentals, as well as a sense of the meaningfulness and quality of the data.

Bungartz is a humorous, understandable often rousing speaker

AI carries a lot of prejudices: That would make models worse. Bungartz: The black box nature of neural networks is certainly a problem - some things work, but you can't really understand why. In science, however, we need to understand and we need reproducibility. That's why I think it really depends on the interaction between classical simulation and AI, that model-based, comprehensible steps are always included. The AI can also find out which of the planned parameters fit. AI alone rarely provides a reliable answer to questions, but it can provide clues and point scientists in the right direction. I see incredible potential in the field of scientific support systems, which can speed up research processes and produce results more quickly.

What do you hope quantum computing will do for simulation? Bungartz: According to the experts, quantum computing cannot yet be fully assessed, but it is important to look at it intensively now. It's a completely different technology, and we only have a vague idea of where it might help. Quantum systems offer advantages in exploiting massive parallelism, allowing different scenarios or simulations to run in parallel. But when this will become commonplace in science is still an open question, and there are certainly people who still have doubts.

What kind of model are you still missing? Or rather, where would you like to contribute your experience in simulation? Bungartz: Things gets exciting when they involve people. Take pedestrian simulation. At first glance, this is not a particularly exciting task, but over the past few years we have been working closely with Professor Gerta Köster from the Munich University of Applied Sciences - she is doing research in exactly this area. The most exciting thing about pedestrian simulation is the human factor. If you want to calculate an evacuation scenario, for example for a bomb threat in a stadium, you will soon reach the end of the line using physical and mechanical perspectives. Social behaviour and psychological components are crucial. What happened in Duisburg in 2010 was often re-enacted and simulated. You want to take psycho-social behaviour into account in such tasks. I think that's a key factor, but we're still at the beginning. People who line up in a room do not spread out evenly, they tend to form small groups. These are effects that can sometimes be explained by hard, physical models and sometimes not. And for panic behaviour, these questions are important. There is still a lot we can do with the human factor in simulation, but it is often not clear how to do it.

Do you at least have an idea of how to express psycho-social behaviour mathematically? Bungartz: First of all, you try to describe it through parameters. In an evacuation scenario, groups try to stay together. Parents won't leave a small child behind, of course. But young people who went to a football match together will also try to stay together. Social bonding is incredibly strong. This can be described in terms of parameters - something similar can be found in chemistry, for example: some molecules are rarely far apart. Parameters are a simplification, but they offer a way of formalising psycho-social phenomena and putting them into an equation. The same is true here: This is a start - there is still a lot to explore. (interview: S. Vieser)