Once every two years, more than a thousand scientists and practitioners in the field of computing gather at the SIAM conference on computational science and engineering (CSE). The 2023 edition in Amsterdam once again offered a showcase of everything that is going on in the field, from software management, digital twins, and machine learning to impressive applications. Here are four takeaways that might be relevant for anyone who is working on computational applications.
1. Managing open-source software is not trivial
About 15% of a large computing code needs to be modified annually, which means that for every 66k lines of code, you need a full-time developer, even apart from new development. That was the statement from Luca Heltai in his keynote speech.
Luca is an associate professor of numerical analysis at SISSA, an international school for advanced study in Triest in Italy. He is a driving force behind the Deal II software, which grew from a small research code of 124k lines of code in 2000 to a huge code base of 1.5M lines of code in 2022. A simple calculation shows that this code needs more than 22 full-time developers just to keep the code base in order.
This would perhaps not be such a big deal if the code was not an open-source development, driven almost entirely by voluntary developers. How do you create and sustain such a community of volunteer developers? Mr. Heltai gladly shared his insights at the SIAM conference in Amsterdam.
One is that you need to have your documentation in order. If developers cannot find their way around the code and do not understand the ideas and rules behind it, the community will not function. In general, setting up all possible forms of support for developers is necessary. In the Deal II software, there is one line of documentation for every 1.6 lines of code. And by the way: for every line of code there is one line of test code. This shows an aspect of software that is often underestimated: the source code is only half (or less) of the business.
Mr. Heltai had other interesting experiences to share, but the audience will probably mostly remember the impressive demonstrations of the Deal II software. The video of an emerging rift in the earth’s surface due to continental drift wowed the auditorium. It made the comprehensive simulation of the heart, which Mr. Heltai also showed, seem almost simple.
2. Digital twins spur a lot of development in computational science
Not surprisingly, digital twins were everywhere at the SIAM CSE conference. Many sessions were dedicated to the subject, but as many dealt with underlying technology like high-performance computing, reduced-order modeling, and data assimilation.
With respect to high-performance computing, much interest these days goes to GPUs, which are quickly becoming, or already are, the standard platform for large computations. The traditional bottleneck of CPU-GPU communication is still a challenge in many real-world applications. But so is the “Babylonian Language Confusion”, as Karl Rupp from the Technical University in Vienna called it. This refers to the various platforms for programming GPUs like CUDA, OpenCL, HIP and OKL/OCCA. With all these different platforms it is sometimes hard to transfer a solution from one environment to the other.
In the field of data assimilation, interesting work was presented by Konstantinos Tatsis from ETH Zurich about simultaneously estimating the state of a system, its parameters and its inputs, and all that from noisy output measurements. The principle behind his work is to build a bank of filters, where each filter is associated with a reduced-order model of a particular instance of the parameters. These filters are combined with weights that are estimated from the data. The entire framework involves some very serious mathematics, but the basic idea seems plausible.
Even as research is still in full swing, the industry is quick to catch on to digital twins as there are many attractive business cases. A company like Siemens is going all in on this emerging technology, as was exemplified by the talk from Herman Van der Auweraer from Siemens and KU Leuven in Belgium. He discussed the principles underlying the toolset that Siemens has developed and showed several interesting examples. One of these is a mechanical test of a wind turbine blade, where augmented reality is used to overlay a stress image over the image of the vibrating blade in real time. Thus, the engineers can actually see the computed stresses as the blade moves.
Even if the talk by Van der Auweraer, at the SIAM CSE conference, had something of a promotional talk for the Siemens tool set, it is still impressive to see what is already there. But it would be better if there was more standardization in the field of digital twins. An effort in that direction was presented in the same session by Oliver Barrowclough from SINTEF, an independent research organization in Norway. He described the ISO 24247 Digital Twin Framework for Manufacturing in the context of additive manufacturing. As an example, he showed how sensors in the production process are used to create a digital twin representation of the manufactured object. This can then be used for virtual testing and quality control, taking away the need to create extra physical instances of the objects only for quality assessment.
3. Machine learning offers new directions in computational science.
Next to digital twins, the other hot topic at the SIAM CSE conference was machine learning. It offers a powerful new toolset to compute things that could hardly be computed before. Weinan E from Beijing University, one of the keynote speakers, illustrated this with examples of high-dimensional systems that occur in quantum mechanics and dynamic programming. Traditional methods for these systems suffer from the curse of dimensionality, but machine learning does not. Mr. Weinan demonstrated this with the results of a simulation of over 100 million copper atoms with the DeepMD-kit. It apparently uses 27.360 GPUs on the Summit supercomputer, so it is still quite a computation. But at least it can be done.
A completely other use of machine learning in computational science is to integrate various sources of information. Machine learning sheds rather new light on what was traditionally known as data fusion. For example, a picture of a dog with a red ball and the associated text label (“a dog with a red ball”) are in fact two sources of information (two “modalities”) that can be combined. Madalina Fiterau from the University of Massachusetts presented a particularly clever way to combine information from different modalities like text, time series, and images. She uses an iterative approach that makes sure that no information is lost in the process. Her method seems very useful for instance for financial data, which is typically multi-modal. It will be interesting to see how traditional data-fusion methods and machine learning approaches like these can further reinforce each other.
4. Applications are crying out for even better computational methods
Although there were probably many attendees that were interested in the research as such, there were certainly also many who came to see what all this science can do for their application. The presenters in the public event on Wednesday evening were of this latter kind. Their talks offered a showcase of topics that are crying out for ever better computational methods. For those who missed it, a recording of the public event is available on Facebook.
This public event was intended to show a non-expert audience the wonderful things mathematics can do for societal challenges. Admittedly, not every talk was easy to understand for non-mathematicians. But even if they might not understand all of it, the public could still enjoy impressive simulations and anecdotes. For example, Luke Bennets from the University of Adelaide showed an impressive video of the flow from the continental shelf of Antarctica. And many will remember the casual remark from Bert Zwart from the Center for Mathematics and Computer Science in Amsterdam, about power grids in England, stating that as long as London is ok, blackouts are not considered problematic there. This illustrates that applications may involve priorities that have nothing to do with perfecting the computations.
Even so, the public event was a strong reminder of the relevance of the field. Computational Science and Engineering is indispensable for addressing the challenges that humanity faces. From pathogens to climate change to the energy transition. The world is fortunate that science is moving fast, and this conference is a testament to that.
If you like to more know about any of the topics that are discussed in this blog, feel free to contact us. We’ll be happy to share our knowledge and experience with you.