While many studies are based on historical data, the research of Rebekah Baglini is an excellent example of supercomputing applied to recent data in the humanities.   

She employs supercomputing in her current projects involving the collection, processing, and annotation of large-scale media data from traditional and social media sources. By examining this diverse range of data, Rebekah Baglini investigates causal inference and causal reasoning from a linguistic perspective. Her research involves the application of semantic model theory and computational methods to uncover insights in linguistics.

“I aim to develop computationally assisted methods to identify trends in the discursive and informational landscape around topics concerning media dynamics, public health and science communication, crisis and risk messaging, as well as the emergence of mis- and dis-information”. 

Rebekah Baglini, Associate Professor in Linguistics, Aarhus University

In addition to her linguistic investigations, Rebekah Baglini also strives to enhance the existing computational language models for multilingual natural language processing (NLP), with a particular focus on under-resourced languages.   

Read the full article about Rebekah Baglini’s project on the DeiC Interactive HPC website.

PI: Associate Professor, Antonio Rago, the Department of Mathematics and Computer Science, SDU

The following is an excerpt from an article published on the eScience Center’s website in August 2023: SDU researcher receives record-breaking HPC grant

Antonio Rago’s field of research is particle physics – more specifically lattice quantum chromodynamics (QCD), a field known for heavy utilization of computational resources.

To use an analogy, Antonio Rago explains that his method could be related to the generation and analysis of a painting, e.g. of the Mona Lisa, where the task is to perform some type of measurement, e.g. of the size of the nostril of Mona Lisa. If the pixels of the painting are too rough, you cannot see the Mona Lisa, let alone measure the nostril. You need a resolution of e.g. a few millimetres. 

Now imagine that in the same painting of Mona Lisa, with a resolution of a few millimetres, you also need to include Mont Blanc in its original size. And you need to develop a type of measurement that is extremely precise and relevant both for the nostril of Mona Lisa and Mont Blanc. This may provide some insight into the difficulties that Antonio Rago and his research group are facing – the particles that they need to measure can namely be extremely different in size and weight. 

“You have to imagine that I want to deal with a matrix of 10²⁰ elements. It is a gigantic amount of data and I want to act on this matrix within some very specific precision for every single element. So you really need a gigantic supercomputer and to have a very precise understanding both of the algorithmic properties and of the hardware, because you cannot afford to lose any information. It’s a terrible tasks,” Antonio Rago admits.

Searching for the origin of asymmetry

The terrible task could be worth the trouble in the end, however, because it might lead us to a better understanding of such big open questions as: why are we alive?

In the universe, there is unbalance between matter and anti-matter. Matter is essentially everything we see around us – from the smallest grain of sand to large planets in the universe. Anti-matter is the opposite, i.e. what is lacking, or matter that is reversed under charge. Since the Big Bang, there has been a competition between matter and anti-matter. They were once equal in size. But some mechanism, which is still unknown to scientists, has resulted in an overwhelming victory of matter over anti-matter, leading to the universe that we know today, which contains comparatively little anti-matter.

The search for this mechanism is exactly where Antonio Rago and his research group believe that they can contribute with a new tool called ”master field simulation”:

“The lattice approach – i.e. the measurements done via simulations on a computer – has somehow reached its saturation. It can not go any more precise than it is now. We are pretty confident that the master fied simulation will allow for a new era for this measurement,” says Antonio Rago.

The project “Master field – infinite volume lattice QCD” has been awarded 120 million CPU core hours on the European supercomputer, LUMI. This is the grant is the largest that has ever been given for HPC to any researcher in Denmark. 

To overcome the challenges of limited data from ancient societies, researchers have started utilizing Agent-based model (ABM) sometimes enabled by supercomputing. ABMs are computational models that simulate the behaviour and interactions of individual entities, known as agents, within a specified environment or system. Each agent in the model is typically programmed with a set of rules or algorithms that control its behaviour, decision-making processes, and interactions with other agents and the environment.

ABM is a valuable tool in archaeology that allows us to simulate and analyse the behaviours and interactions of individuals or groups in past societies, and the use of ABM allows comparison of the model against real archaeological data.

Assistant Professor Iza Romanowska

In one of Iza Romanowska’s studies, agent-based modelling (ABM) made it possible for her and her colleagues to explore the Roman economy in the context of long-distance trade, using ceramic tableware to understand the distribution patterns and buying strategies of traders in the Eastern Mediterranean between 200 BC and AD 300.  

You can read the full article about Iza Romanowska’s project on the DeiC Interactive HPC website.

Katrine Frøkjær Baunvig, head of the Grundtvig Center at Aarhus University, has used supercomputing as a methodological approach, and it has led her to non-trivial conclusions that significantly impact our understanding of of 19th-century nation builder and prominent pastor N.F.S. Grundtvig ‘s vast body of works and his immense influence on Danish culture.  

Go to the DeiC Interactive HPC wesbite to read more.  

Co-Pi: Associate Prof. Benjamin Jäger, Department of Mathematics and Computer Science, SDU 

Though supercomputers form the key basis of postdoc Tobias Tsang’s research, UCloud has been a valuable, complementary tool for him and his colleagues and will most likely continue to be so in future work as well.

“I don’t think UCloud will ever be the only resource we use. But this is also the design of it; UCloud is not meant to be a huge machine, it is meant to be an available resource that is easy to use and that gives you a playground to set up things really from scratch where you can test things out and run smaller jobs and analyses. In that sense, it is quite complementary to a lot of the things we normally work with. For exploratory studies and for code testing, UCloud will definitely remain very useful.”

Postdoc Tobias Tsang

At one specific project, “One Flavour QCD as an analogue computer for SUSY”^[1], done at SDU as a collaboration between CP3-Origins ( Centre for Cosmology and Particle Physics Phenomenology) and IMADA (Department of Mathematics and Computer Science), the vast majority of samples were generated on UCloud, and a significant amount of data production and measurements were also carried out on there. UCloud needs, however, to be considered a part of a whole, according to Tobias:

“It is not that one particular machine made it possible; we would otherwise have found another machine to run it on. But UCloud provided us with a nice set up where we could just use local resources without having to go through big grant applications to get computer time.”

Postdoc Tobias Tsang

Tobias stresses the interaction with the UCloud front office as a major benefit that has helped the research group significantly, especially compared to other clusters with a much longer response time:

“One of the nice things with UCloud as a general system is that every time something didn’t work, we got a really quick email back. Any questions we raised were answered quickly, so it was never something that kept us stuck for weeks or months – typically things were resolved in a very timely time scale. And things that we actively suggested as nice features or things that we thought were missing on UCloud were likewise addressed.“

Postdoc Tobias Tsang

^[1] Della Morte, Jaeger, Sannino, Tsang and Ziegler, “One Flavour QCD as an analogue computer for SUSY”, PoS LATTICE2021 (2022) 225, https://doi.org/10.22323/1.396.0225

Course PIs: Mikkel Heide Schierup, Aarhus University, and Stig Uggerhøj Andersen, Aarhus University.

As part of the project National Health Data Science Sandbox for Teaching and Training, Samuele Soraggi and researchers at Aarhus University ran a course on UCloud called Introduction to Next Generation Sequencing Data. 

“UCloud is pretty ideal for organizing pedagogic material because it is interactive. The user support team at the eScience Center helped us so we basically had an app running and starting the packages so the users would not have to install anything and could just straight use the course material interactively,” says Samuele Soraggi. 

The course has been published as an app on UCloud, Genomics Sandbox, which can be used by other researchers, organizing similar courses. A population genomics and an introductory GWAS course are on their way to the Genomics Sandbox in the coming months.

“I think it’s very nice that the course participants did not have to worry about packages and installation. They could also upload their own datasets and try out the same analysis for their research projects. That was a plus we did not think about, because we thought mostly about the tutorial for the course,” says Samuele Soraggi.

If you would like to host a similar course on UCloud, please contact support@escience.sdu.dk. If your course is related to health science, the team behind the National Health Data Science Sandbox can also help you at NHDS_sandbox@sund.ku.

The project Detecting text reuse in H.C. Andersen’s work is part of a larger publication project, Hans Christian Andersen’s Fairy Tales and Stories – the digital manuscript edition, led by Associate Professor Ane Grum-Schwensen, which aims to digitalize and publish all the preserved manuscripts of Hans Christian Andersen in an online, genetic edition (if you are interested in knowing more about the digital edition, it is described in more detail at http://andersen.sdu.dk/ms. The updated description is in Danish, but an older English version is available at http://beta.auh.sdu.dk/en/.) 

In 2019, senior researcher Ejnar Stig Askgaard from Odense City Museums began comparing Hans Christian Andersen’s notes, written between approximately 1833 – 1875, with the 162 fairy tales, novels and autobiographies. This had led to the discovery that Hans Christian Andersen liked to use symbols such as cross marks or deletions in his notes to indicate that the note had been reused in his fairytales. 

For Detecting text reuse in H.C. Andersen’s work, Berg wanted to find out where each note had been reused. Earlier research had managed to manually identify where 278 notes had been reused in Hans Christian Andersen’s published work, but this had been a time-consuming effort, taking many months of work.

As 861 of the notes had been digitalized in addition to Hans Christian Andersen’s published work, Berg was able to apply digital methods to solve his problem. He contacted Zhiru Sun, Assistant Professor at the Department of Design and Communication at SDU, who used a method called Natural Language Processing to find similarities between the notes and Hans Christian Anderson’s work. Using the Python application on UCloud, this method generated a number of tables, which indicated how similar a specific note is to a specific fairytale.

“It only took me around 8 hours to generate these tables and find a good indication of where all the 861 digitalized notes had been reused,” Sun explains. 

In the tables above, each note has received a score from -1 to 1. The closer the score is to 1, the more similar is the note to the fairytale and vice versa. Note_61, e.g., where the low score indicates that it is very different from all fairytales, is a shopping list.

You can read a more detailed interview with Zhiru Sun here.

“I started my work in SDU as Lundbeck Prof. of Neuroimaging from the beginning of this year. My job is to setup a Neuroimaging Research Team in SDU. To do this computational infrastructure is essential. This usually is costly and requires considerable expertise. The eScience team provided a superb service to my team and together we developed Neuroimaging Analysis Platform which is conveniently available to all UCloud users on site or remotely for my research collaborators. The platform makes the best use of considerable CPU and GPUs available in UCloud. This platform will only get better as more people use it and provide feedback. Without such infrastructure and expertise available at eScience, our research would be impossible.

We analysed data from AHBA and ENIGMA datasets to interrogate the transcription-imaging co-locations in deletion syndrome (D22), autism spectrum disorder (ASD), attention deficit hyperactivity disorder (ADHD), high risk psychosis (HRP), schizophrenia (SCz), bipolar disorder (BD), major depressive disorder (MDD) and obsessive-compulsive disorder (OCD). This study included MRI data from 30,000 subjects. Complex mathematical and statistical analysis was used to find combination of genes (among 20,000 genes) that are maximally covaried with brain MRI changes. For each PLS, bootstrapping (10,000 times) and spin permutation (10,000 times) were performed to avoid random associations. Such analysis would have taken (if at all possible) over 6 months on a high-power personal computer, but we used Neuroimaging Platform on UCloud and managed to seamlessly reach the result in two weeks.” 

Mojtaba Zarei, Clinical Prof.

Figure 1. Spatial association between gene expression and brain atrophy across psychiatric disorders.

Figure 2. Gene expression similarity across different psychiatric disorders.

Prof. WSR Oliver Baumann collaborates with two American colleagues to study resource allocation in hierarchical organizations, and how resources should be distributed between more well-known approaches and more innovative/risky ideas, depending on the structure of the organization.

The following figure illustrates two stylized organizational structures. In the model, a “senior manager” needs to decide how to allocate (limited) resources among his/her “middle managers”, who in turn need to decide how to allocate the resources they received among a set of uncertain alternatives (e.g., projects exploiting different technologies whose value is not fully understood). This set-up captures a typical learning process in organizations. Middle managers are specialists in the sense that they are dealing with concrete projects. By running a project, the middle managers can learn about the value of the underlying technology. Senior managers, on the other hand, are more generalists that learn about their middle managers, i.e., about the aggregate performance of the projects for which a particular middle manager is responsible.

Baumann and his research group develop simulation models that involve Monte Carlo methods. This approach relies on repeated random sampling to obtain numerical results and predict the probability of different outcomes. Therefore, it also creates a need for running many replications of the same model to obtain good approximations. This is where HPC resources come into the game. 

“We can do things much faster, and we can do more than we ever could with our own machines. Computation is also cheaper and more easily available, and this means we get to experiment more.”   

Prof. WSR Oliver Baumann

The HPC resources that Baumann and his group use are distributed between Ohio State University, where his two colleagues work, and UCloud at the DeiC Interactive HPC facility at SDU.

We speculate that different locations in the DNA are exposed to a varying degree of evolutionary pressure which differs between populations based on the genetic background, the environment and the geographic locations. It is already known that various population groups have different rates of SNPs (single nucleotide polymorphisms) and there are also SNPs which are rare globally, but occur frequently in a specific population. It is likely that this is not only the case for single SNPs, but true for entire regions in the DNA with important implications to the susceptibility of entire populations towards certain diseases or conditions – all of which could explain the different prevalence of disease subtypes between populations. Besides environmental factors, different local distributions of the SNP densities in one sub-population might lead to different consequences, as in the patient group, where the same SNP is in the evolutionarily more conserved region.

Therefore, we are developing a novel method for large-scale analysis of genetic data unravelling statistically significant distributions of genetic traits and modifications in the DNA between different populations.

The increasing number of human genomes allows for the in-depth analysis of the evolutionary conservation. More and more consortia are formed and produce a tremendous wealth of data which is rapidly made available to the public, for example the 1000 genomes projects, several national initiatives like Denmark or Great Britain which seek to sequence over 100,000 genomes.

Analyzing the entire genome as well as using the available data from different consortia requires a lot of computational power therefore, all the processing, data handling and calculations used in the project are performed on ABACUS2.0, located at SDU in Odense, Denmark.