Programme

Friday 10 November 2023
Time
Presentation title
09:00 - 09:45
Registration and Coffee
09:45 - 10:00
Opening and Welcome – Johan Mentink (RU) and Kees Vuik (TUD)
Johan Mentink visit profile
10:00 - 10:30
Keynote: Speeding up calculations for water management
Location: Plenary Room - Juliana Congreszaal
Albrecht Weerts - WUR & Deltares visit profile
Presentation category: Multiscale Modeling
Presentation summary
Speeding up calculations for water management
10:30 - 11:00
Keynote: From multiscale modeling of knee osteoarthritis to building the virtual human twin
Location: Plenary Room - Juliana Congreszaal
Liesbet Geris - University of Liège, KU Leuven & Virtual Physiological Human Institute, Belgium visit profile
Presentation category: Multiscale Modeling
Presentation summary
From multiscale modeling of knee osteoarthritis to building the virtual human twin
11:00 - 11:30
Coffee Break
11:30 - 12:00
Keynote: Application of deep learning to drug discovery: limitations and opportunities
Location: Plenary Room - Juliana Congreszaal
Francesca Grisoni - Eindhoven University of Technology, Dept. Biomedical Engineering, Institute for Complex Molecular Systems (ICMS) and Eindhoven AI Systems Institute (EAISI) visit profile
Presentation category: Data-driven methods
Presentation summary
A discussion on how deep learning can be applied to drug discovery, and the current limitations and opportunities.
Abstract details
Deep learning is fueling science and technology, e.g., to predict protein structure, advance mathematics and investigate galaxies. Computer-aided drug discovery has particularly befitted from the advent of deep learning. This talk will illustrate some successful applications of deep learning for de novo molecule design and molecular property prediction, at the interface between theory and wet-lab experiments. Moreover, it will provide a personal perspective on current limitations and future opportunities for deep learning in medicinal and organic chemistry, to accelerate molecule discovery and chemical space exploration.
12:00 - 13:30
Lunch
12:00 - 13:30
Poster presentations
13:30 - 15:00
Data-driven methods
Location: Room: Juliana 1
Gennadii Donchyts - Google
Bernadette Mohr - University Of Amsterdam
Tim Offermans - FrieslandCampina
Jeremie Vandenplas - Wageningen University & Research
Clelia de Mulatier - University of Amsterdam, Computational Science Lab
Session details
13:30 - 14:00
Google Earth Engine as a data driven platform for sustainable decision-making - 1
Gennadii Donchyts - Google visit profile
Presentation category: Data-driven methods
Abstract details
Google Earth Engine is a cloud-based platform for planetary-scale analysis of geospatial data. It provides a powerful set of tools and data for monitoring, understanding, and responding to environmental change. Earth Engine was founded in 2009 to bring the power of datacenter-scale computing to bear on global challenges facing the Earth and its inhabitants. It has been the groundbreaking platform for planetary scale remote sensing analysis since its inception. In this talk, I will provide an overview of the platform and will discuss how Google Earth Engine is used to support sustainability initiatives. I will present examples of how Google Earth Engine has been used to track deforestation, improve agricultural practices, map and monitor water changes, and assess the impact of climate change and disaster events. I will also discuss the challenges and opportunities for using Google Earth Engine to support sustainability initiatives in research, commercial, and operational contexts.
14:00 - 14:15
Condensed-phase molecular representation to link structure and thermodynamics in molecular dynamics -
Bernadette Mohr - University Of Amsterdam visit profile
Presentation category:
Abstract details
Molecular design requires systematic and broadly applicable methods to extract structure-property relationships. The focus of this study is on learning thermodynamic properties from molecular-liquid simulations. The methodology relies on an atomic representation originally developed for electronic properties: the Spectrum of London and Axilrod-Teller-Muto representation (SLATM). SLATM's expansion in one-, two-, and three-body interactions makes it amenable to probing structural ordering in molecular liquids. We show that such representation encodes enough critical information to permit the learning of thermodynamic properties via linear methods. We demonstrate our approach on the preferential insertion of small solute molecules toward cardiolipin membranes and monitor selectivity against a similar lipid. Our analysis reveals simple, interpretable relationships between two- and three-body interactions and selectivity, identifies key interactions to build optimal prototypical solutes, and charts a two-dimensional projection that displays clearly separated basins. The methodology is generally applicable to a variety of thermodynamic properties.
14:15 - 14:30
Process expert knowledge is essential in creating value from data-driven industrial soft sensors -
Tim Offermans - FrieslandCampina visit profile
Presentation category:
Abstract details
Industry 5.0 aims to put human operators back at the center of digital process automation, going beyond the Industry 4.0 goal of autonomous processes that require no human intervention. This work shows an integrated approach to using contemporary data-driven technology to obtain industrial soft sensors that are systematically enriched with expert knowledge in a quantifiable and demonstrable manner. We develop, investigate, and compare data- and expert-driven approaches for selecting process variables for real-time product quality prediction for three parallel-operated dairy processing lines. We find that expert-driven variable selections outperform well-established data-driven selections. However, the overall largest value is obtained when data-driven algorithms are used to perform selection on a series of variables that were pre-curated by process experts. This seems most likely because experts can more easily pick up on variation over time due to instability of process variables, which is difficult for data-driven methods to do.
14:30 - 14:45
Computational challenges for solving large-scale genomic mixed model equations -
Jeremie Vandenplas - Wageningen University & Research visit profile
Presentation category:
Abstract details
In livestock production systems, evaluating the genetic merit of animals is a key process to stepwise improve a population for some characteristics of interest in each next generation. Such evaluations of the genetic merits of all animals are obtained by solving systems of equations derived from a statistical linear mixed model that analyses simultaneously performances and genealogical information of animals with or without known DNA-profiles. Those DNA-profiles consist of genotypes that only include values of 0, 1, 2, or missing values. The main computational challenges of these systems are that they contain hundreds of millions of equations, and that their coefficient matrices are ill-conditioned and mainly sparse, except for a dense block associated with up to a few millions of equations related to the genomic information. We will first present our matrix-free approaches for solving efficiently these systems with a two-level preconditioned conjugate gradient approach. Additionally, for coping with growing genotype datasets, we will present computational advantages of bit-level algorithms developed for compressing, storing, and multiplying a genotype matrix with a double-precision matrix. Such algorithms require 32 times less RAM and are at least twice as fast as the same multiplication with the Intel MKL DGEMM procedure on CPU architectures. The compression of the genotype matrix also allows the use of Nvidia GPUs, resulting in an additional speed-up of a factor up to 20.
14:45 - 15:00
Robust detection of high-order community structures in noisy binary data -
Clelia de Mulatier - University of Amsterdam, Computational Science Lab visit profile
Presentation category:
Abstract details
Uncovering groups, or “communities”, of highly correlated variables in noisy data is challenging yet crucial to understanding emergent phenomena in complex systems like the brain, health, or social systems. Recently, there has been a surge of interest in the impact of high-order interactions on the behavior of complex systems. So far, existing techniques for identifying communities in data rely solely on the pairwise correlation patterns of the data. In contrast, we introduce a novel approach that takes into account all higher-order data patterns in identifying communities. The method consists of performing Bayesian model selection with a new family of probabilistic models, known as Minimally Complex Models (MCMs), which have a clear interpretation in terms of community structure. Our method identifies optimal communities by maximizing the evidence of MCMs. We discuss two algorithms for performing such optimization. We tested our method against modularity optimization on noisy data generated from benchmark graphs and hyper-graphs that have built-in community structure with varying overlaps. We show that our approach can recover robust community structures hidden in noisy data, even in undersampled cases. Finally, our analysis reveals a specific region of the phase diagram where the pairwise analysis fails to recover the true communities, but our high-order analysis can. This highlights the effectiveness of our approach in capturing complex high-order community structures.
13:30 - 15:00
Energy-efficient computing - 2
Location: Room: Juliana 2
Benjamin Czaja - SURF BV
Alessio Sclocco - Netherlands eScience Center
Matthias Möller - Delft University of Technology / Department of Applied Mathematics
Fang Fang - FF Quant Advisory
Arne Wulff - TU Delft
Session details
13:30 - 13:55
Energy aware computing on the Dutch National Supercomputer Snellius -
Benjamin Czaja - SURF BV visit profile
Presentation category: Energy-efficient computing
Presentation summary
The energy usage of computing infrastructure is becoming an increasingly prominent topic for computing systems design and usage. It has been estimated that the electrical energy foot print of data centers accounts for 1% of the global energy footprint, and has been projected to grow to 10% by 2030 [1]. At the Dutch National Supercomputer Snellius we are actively studying and promoting energy aware computing in terms of exploring new architecture design as well as carrying out energy analysis of research software [2].
Abstract details
In this talk we quantify the electrical energy usage of Snellius during production. Specifically, we monitor and detail the energy usage, and characteristics, of research software that is actively being deployed on Snellius. Once the energy and performance characteristics of an application are understood we highlight end user tools [3] that can be employed to minimize the energy footprint of an application either through optimization of resource utilization or cpu frequency throttling which decreases energy usage while limiting the degradation to performance. The motivation of this work is to provide an overview on how researchers can optimize the use of computing systems and application development for the purpose of energy efficiency via the use of simple tools. We believe that this work will lead to a more informed research community about the energy usage of their applications and increase the understanding energy usage of ICT infrastructure for the future. [1] Jones, Nicola. "How to stop data centres from gobbling up the world’s electricity." Nature 561.7722 (2018): 163-166. [2] Dolas, Sagar, Ana Verbanescu, and Benjamin Czaja. "Making Scientific Research on Dutch National Supercomputer Energy Efficient." ERCIM News 2022.131 (2022). [3] J. Corbalan, L. Brochard, “EAR - Energy Management framework for Supercomputers”. https://www.bsc.es/sites/default/files/public/bscw2/content/software-app/technical-documentation/ear.pdf
13:55 - 14:15
Optimizing your code for energy-efficiency -
Alessio Sclocco - Netherlands eScience Center visit profile
Presentation category: Energy-efficient computing
Presentation summary
In this talk, we will highlight tools and techniques that we use at the Netherlands eScience Center to improve the energy efficiency of the code we work on, and provide examples of the results we have achieved in recent years.
Abstract details
From large-scale experiments and simulations, to analyzing and visualizing data, the use and development of software is nowadays an integral part of the job routine for many scientists. Running this software, no matter if on a laptop, the cloud, or a supercomputer, has a cost in terms of energy. While in the past this cost was mostly a concern for the provider, the rising cost of electricity combined with the increased power requirements of supercomputers, now in the megawatt range, have resulted in some providers passing this cost on to the users. This fact, paired with the ongoing effort to reduce the environmental impact of science and its carbon footprint, is forcing the scientific community to reduce the energy consumption of their software. Traditionally, optimizing time to solution was considered the optimal way to also reduce energy consumption. However, new results suggest that this is not always the case, and that optimizing directly for energy efficiency can reduce overall energy consumption without compromising performance. In this talk, we will highlight tools and techniques that we use at the Netherlands eScience Center to improve the energy efficiency of the code we work on, and provide examples of the results we have achieved in recent years.
14:15 - 14:30
Quantum Computational Fluid Dynamics -
Matthias Möller - Delft University of Technology / Department of Applied Mathematics visit profile
Presentation category:
Abstract details
Computational Fluid Dynamics (CFD) has been pushing HPC systems to their limits since decades. Despite all advancements in numerical methods and HPC technologies, fully resolved turbulent flow simulation at large scale remains out of reach with traditional computer architectures due to unsatisfiable resource and time requirements. Last but not least, modern HPC systems consume an enormous amount of energy for operation and cooling. In this talk we discuss opportunities to complement classical HPC systems with quantum computers to push CFD to a next level and also mitigate the enormous energy hunger of classical HPC systems. In particular, we present a quantum lattice Boltzmann method that can be run start-to-end on a (future) fault-tolerant quantum computer. Next to discussing data encodings and quantum primitives for the three core steps of any lattice Boltzmann solver - streaming, collision and reflection - we will shed some light on their computational complexity and show preliminary results computed on a quantum computer simulator.
14:30 - 14:45
A Novel Algorithm for Computing Multi-dimensional Expectations By the Dimension-Reduced COS Method -
Fang Fang - FF Quant Advisory visit profile
Presentation category: Data-driven methods
Abstract details
A wide range of problems faced by practitioners in the financial industry are in essence multivariate expectation problems. Typical examples of such include the valuation of multi-asset options and risk quantification on portfolio level. Very often there exist no analytical solutions for those expectation problems and we thus have to rely on numerical methods. In this paper, we tackle these problems from a novel angle. The idea is to directly solve the characteristic function (ch.f.) of the combined dynamics of the driving random variables, with which the targeted expectation can be easily recovered. The ch.f. itself is again defined by a multivariate expectation, which is approximated by a dimension-reduced Fourier-cosine series expansion of the joint density function of the driving random variables, based Canonical Polyadic Decomposition (CPD). Hence, we name this method "dimension-reduced COS method". We illustrate its application using XVA pricing as an example. The error convergence is proved mathematically and the performance is tested to be much superior to straight-forward numerical integration or Monte Carlo simulation.
14:45 - 15:00
Quantum-enhanced AI for sustainable Materials and Structural design in aerospace (QAIMS) -
Arne Wulff - TU Delft visit profile
Presentation category:
Abstract details
Quantum computing and machine learning are emerging as transformative tools capable of addressing complex, computationally intensive challenges. Our research group investigates how these tools can be leveraged in materials and structural design for aerospace applications. Comprising four PhDs, we focus on two primary areas: the optimization of laminates and molecular simulations of high entropy materials. Both domains are geared towards advancing innovative material and structural solutions for aerospace needs. This presentation will provide a comprehensive overview of our research group's projects and objectives.
13:30 - 15:00
Uncertainty Quantification - 3
Location: Room: Juliana 3
Matias Valdenegro Toro - University of Groningen
Harshit Bansal - Eindhoven University of Technology
Wouter van Harten - Radboud University
Laura Scarabosio - Radboud University
Daan Crommelin - CWI / University of Amsterdam
Session details
13:30 - 13:45
Introduction to Uncertainty in Machine Learning with Disentangling -
Matias Valdenegro Toro - University of Groningen visit profile
Presentation category: Uncertainty Quantification
Presentation summary
Uncertainty in Machine Learning models is required to obtain users' trust and signal when predictions are likely to be correct or should be discarded. In this presentation we will cover basic details of how uncertainty is estimated in machine learning models, with focus on having separate predictions for data and model uncertainty, often called uncertainty disentanglement.
Abstract details
Uncertainty in Machine Learning models is required to obtain users' trust and signal when predictions are likely to be correct or should be discarded. In this presentation we will cover basic details of how uncertainty is estimated in machine learning models, with focus on having separate predictions for data and model uncertainty, often called uncertainty disentanglement. This allows modelling of different sources for uncertainty, like separating noise in the labels from the uncertainty of the model itself, impacting the quality of decisions made with uncertainty values. We will present regression results to separate data and model uncertainty, based on results from our previous research, in particular the paper "A Deeper Look into Aleatoric and Epistemic Uncertainty Disentanglement" which was published at CVPR Workshops ( https://arxiv.org/abs/2204.09308 ).
13:45 - 14:00
Bayesian optimal experimental design in the presence of model error -
Harshit Bansal - Eindhoven University of Technology visit profile
Presentation category: Uncertainty Quantification
Abstract details
Solving Bayesian inverse problems and performing a Bayesian optimal experimental design (OED) are ubiquitous in various scientific domains, e.g., geothermal applications. To this end, several computationally expensive forward problems need to be solved. In order to accelerate the inverse estimation or optimal sensor placement (or OED), surrogate models have come to the fore. However, due to the incurred approximation error, usage of (the best) surrogate model on a downstream task, such as inference, leads to a posterior distribution that differs from the underlying true one and, hence, an incorrect estimate. The following methodologies can be adopted to tackle the aforementioned issue: (i) For some fixed sensor placement (or OED), correct the surrogate model used in the Bayesian setting, (ii) Develop an approximation error (or model discrepancy) aware OED algorithm, i.e., define or construct observation operators and place sensors while leveraging the established error bounds for varied classes and qualities of surrogate models. Taking inspiration from the recent papers, which develop a robust OED algorithm under prior or noise covariance misspecification, we adopt the latter strategy to propose a framework that is robust to approximation errors in the surrogate models. We finally test our framework on a benchmark advection-diffusion PDE problem using traditional and newly introduced metrics for Bayesian OED under model misspecification.
14:00 - 14:15
Application of Particle-Based Bayesian Inversion to buried medium voltage power cables in the Netherlands -
Wouter van Harten - Radboud University visit profile
Presentation category: Data-driven methods
Presentation summary
In this research, we employ a particle-based Bayesian inversion algorithm to address the practical challenge of obtaining less conservative bounds for the operation of underground, medium-voltage power cables. The greater objective is to work towards a real-time enhancement of the efficiency and performance of underground cables in the Netherlands. In this project, we aim to deduce the thermal properties of the surrounding soil from an observed thermal profile, as well as quantifying the reliability (or uncertainty in) the obtained estimates. We obtained the thermal profile of the cables by in-the-field measurements in the Netherlands, collected in the last few months. This study is being conducted in collaboration with Alliander, a promine
14:15 - 14:30
Bayesian Calibration and Model Comparison for Tumour Cell Dynamics Under Different Environmental Conditions -
Laura Scarabosio - Radboud University visit profile
Presentation category: Uncertainty Quantification
Presentation summary
I will discuss a methodology to calibrate mathematical models for tumour cell dynamics that take into account the effects of their surrounding microenvironment. The adoption of a Bayesian framework allows to quantify the uncertainty in the parameter estimates as well as possible correlations in the parameters. We will also see how the Bayesian framework can be useful to compare different models. I will present calibration results using different in-vitro laboratory data, where the environmental factors are either nutrient deprivation or oxygen level, tissue stiffness and the presence of chemotherapeutic drugs. The analysis of the calibration results will allow to shed some light on the underlying biological mechanisms affecting tumour progression. This is joint work with Sabrina Schoenfeld (TU Munich), Christina Kuttler (TU Munich), Alican Ozkan (Harvard University) and Marissa Nichole Rylander (UT Austin).
Abstract details
I will discuss a methodology to calibrate mathematical models for tumour cell dynamics that take into account the effects of their surrounding microenvironment. The adoption of a Bayesian framework allows to quantify the uncertainty in the parameter estimates as well as possible correlations in the parameters. We will also see how the Bayesian framework can be useful to compare different models. I will present calibration results using different in-vitro laboratory data, where the environmental factors are either nutrient deprivation or oxygen level, tissue stiffness and the presence of chemotherapeutic drugs. The analysis of the calibration results will allow to shed some light on the underlying biological mechanisms affecting tumour progression. This is joint work with Sabrina Schoenfeld (TU Munich), Christina Kuttler (TU Munich), Alican Ozkan (Harvard University) and Marissa Nichole Rylander (UT Austin).
14:30 - 15:00
Uncertainty Quantification Simulation and model uncertainty -
Daan Crommelin - CWI / University of Amsterdam visit profile
Presentation category: Uncertainty Quantification
Abstract details
In computational studies, the uncertainties in simulation outcomes can stem from various sources, such as parameter uncertainty, initial state uncertainty and structural model error. I will briefly discuss some of the methods that have been developed to quantify these uncertainties, and pay special attention to the topic of model error (or model uncertainty). Regarding the latter, even for systems whose underlying governing principles are known, model uncertainty can arise if we are unable to include all relevant system degrees of freedom in our numerical simulations due to computational cost constraints. Notable examples include climate modeling and fluid dynamics, where the effects of the left-out (unresolved) degrees of freedom are represented in the computational model in a simplified manner through so-called parameterizations or model closures. With stochastic methods, we can account for the uncertainty in these representations.
13:30 - 15:00
Machine learning
Location: Room: Juliana 4
Sascha Caron - Radboud University and Nikhef
Benjamin Sanderse - CWI
Alexander Heinlein - Delft University of Technology
Max Kerr Winter - TU Eindhoven
Gijs van den Oord - Netherlands eScience Center
Session details
13:30 - 14:00
Past, Present and Future of AI in High Energy Physics and Astrophysics: From simple models to GPT4 -
Sascha Caron - Radboud University and Nikhef visit profile
Presentation category: Data-driven methods, Energy-efficient computing, Ultrafast Simulations using Machine Learning
Abstract details
In recent years, the application of machine learning (ML) in particle physics has undergone a rapid transformation. In particular, Deep Learning (DL) technologies, which have been gaining momentum since 2015, are redefining the methods for data analysis in the field. In this talk, I will present several practical examples of DL implementation, including anomaly detection, event classification, particle trajectory reconstruction, and automatic astrophysical object identification. Finally, I will discuss the significant impact of the emergence of powerful foundation models such as GPT-4 and anticipate their potential as a catalyst for the next phase of ML development in science (physics). In addition, I will offer suggestions on how we can proactively address and take advantage of this impending change.
14:00 - 14:15
Learning neural closure models for fluid flows -
Benjamin Sanderse - CWI visit profile
Presentation category:
Abstract details
Discovering physics models is an ongoing, fundamental challenge in computational science. In fluid flow problems, this problem is usually known as the “closure problem”, and the art is to discover a “closure model” that represents the effect of the small scales on the large scales. Well-known examples appear in large eddy simulations (LES) and in reduced-order models (ROMs). Recently, it appeared that highly accurate closure models can be constructed by using neural networks. However, integrating the neural network into the physics models (“neural closure models”) is typically prone to numerical instabilities as the training environment does not match the prediction environment. Instead, we investigate learning closure models while they are embedded in a discretized PDE solver by using differentiable programming software. This is a more difficult learning problem and we present several time integration strategies to deal with the adjoint problem. Furthermore, we present a new neural closure model form which allows us to preserve structure, namely kinetic energy conservation, and therefore non-linear stability bounds.
14:14 - 14:30
Decomposing physics-informed neural networks -
Alexander Heinlein - Delft University of Technology visit profile
Presentation category:
Abstract details
Scientific machine learning (SciML) is a rapidly evolving field of research that combines techniques from scientific computing and machine learning. A major branch of SciML is the approximation of the solutions of partial differential equations (PDEs) using neural networks. In classical physics-informed neural networks (PINNs), simple feed-forward neural networks are employed to discretize a PDE. The loss function may include a combination of data (e.g., initial, boundary, and/or measurement data) and the residual of the PDE. Challenging applications, such as multiscale problems, require neural networks with high capacity, and the training is often not robust and may take large numbers of iterations. In this talk, domain decomposition-based network architectures for PINNs using the finite basis physics-informed neural network (FBPINN) approach will be discussed. In particular, the global network function is constructed as a combination of local network functions defined on an overlapping domain decomposition. Similar to classical domain decomposition methods, the one-level method generally lacks scalability, but scalability can be achieved by introducing a multi-level hierarchy of overlapping domain decompositions. The performance of the multi-level FBPINN method will be investigated based on numerical results for several model problems, showing robust convergence for up to 64 subdomains on the finest level and challenging multi-frequency problems.
14:30 - 14:45
How glassy are neural networks? -
Max Kerr Winter - TU Eindhoven visit profile
Presentation category:
Abstract details
Recently there has been significant interest within the statistical physics community in exploring the similarities between glasses and deep neural networks (DNNs), with the hope of applying the extensive body of glass theory to deep learning. Glasses and DNNs share important similarities: they exist within a high dimensional, non-convex (energy or loss) landscape, and obey gradient descent dynamics subject to noise. In this work we study the training dynamics of DNNs in realistic scenarios through the lens of glassy physics. Despite demonstrating the existence of static glass-like phase transitions in the loss landscape of the DNNs, we find the training dynamics are surprisingly dissimilar to glasses. In particular DNNs lack the divergent timescales that define the glass state. We do, however, find a rich variety of training dynamics that we are able to interpret and model with reference to standard systems from statistical physics. We demonstrate the existence of two distinct training regimes, one dominated by noise, and the other dominated by the gradients of the landscape, as well as aging effects and violation of the Stokes-Einstein relation. Our results allow us to infer a detailed picture of DNN loss landscapes, as well as further strengthening the link between statistical physics and deep learning.
14:45 - 15:00
Hyperparameter optimization of neural networks for proton structure analyses -
Gijs van den Oord - Netherlands eScience Center visit profile
Presentation category:
Abstract details
Progress in high-energy physics is critically dependent upon the robust interpretation of the experimental data that is harvested from frontier facilities such as the LHC at CERN in Geneva. Analyzing the massive amounts of data produced in these particle collisions and comparing these to theoretical predictions enables testing fundamental laws of Nature with unprecedented resolution. For this, a robust understanding of the quark and gluon (i.e. partonic) substructure of protons entering LHC collisions is of utmost importance. Due to the fact that this substructure is determined by poorly understood non-perturbative phenomena, it is advisable to constrain it directly from data. Within the NNPDF approach, this fitting procedure is based on machine learning techniques, training multilayer perceptrons. One of the challenges is identifying an optimal neural network architecture that converges rapidly whilst reducing the risk of overfitting. Here we present an improved hyperparameter optimization strategy, based on new figures of merit that depend on the full probability distribution of the results. This is made possible by enabling the parallel training of hundreds of NNPDF networks by means of GPUs. We compare different hyper-optimization figures of merit and examine their impact in the determination of the proton substructure. Our approach is also relevant for related applications of supervised machine learning in high-energy physics, sharing the challenges of robustly identify
13:30 - 15:00
Multiscale modelling
Location: Plenary Room - Juliana Congreszaal
Siewert-Jan Marrink - Univ. of Groningen
Vítor V. Vasconcelos - University of Amsterdam
Ioana Ilie - University of Amsterdam
Lourens Veen - Netherlands eScience Center
Richard Stevens - University of Twente
Session details
13:30 - 14:00
Computational microscopy of the Cell -
Siewert-Jan Marrink - Univ. of Groningen visit profile
Presentation category: Multiscale Modeling
Presentation summary
The ultimate microscope, directed at a biological cell, would reveal the dynamics of all the cell’s components with atomic resolution. In contrast to their real-world counterparts, computational microscopes are currently on the brink of meeting this challenge. In this talk, I will provide the state of the art on the use of coarse-grain molecular dynamics to simulate complex cellular processes, and provide a perspective on our current efforts to model at the whole cell level.
14:00 - 14:15
The influence of social networks in heterogeneous and variable environments -
Vítor V. Vasconcelos - University of Amsterdam visit profile
Presentation category:
Abstract details
Understanding how collective behavior emerges and prevails in populations remains a significant problem in biological and social sciences. The structure of interactions between individuals is known to play an important role and can be studied using complex networks. However, such an approach often builds on the assumption that decisions are to a large degree accurate and individual properties, random exploration, or migration are rare. However, humans face various largely unpredictable life events that are unrelated to their social interactions but, instead, are conditioned by their life histories — including the level of education, age, housing conditions, income, or wealth---that affect how they respond to their natural, physical, and social environment. Further, individual responses to their environment, at any point in time, are unique to their context and, consequently, heterogeneous. While for a single individual, the influence of others is only a relatively small contribution to their behavior or behavioral change, when embedded in social networks of different topologies and/or assortments of individual characteristics can result in unexpected deviations in the aggregated/macroscopic tendencies in behavior. In this research, we ask how properties of the social networks and the distribution of heterogeneous agents over them affect the aggregated choices of the population and how to nudge them.
14:15 - 14:30
Effects of biological therapeutics on the dynamics of the cellular prion protein -
Ioana Ilie - University of Amsterdam visit profile
Presentation category:
Abstract details
Prion diseases are associated with the conversion of the cellular prion protein (PrPC) into a pathogenic conformer. A proposed therapeutic approach to avoid the pathogenic transformation is to develop monoclonal antibodies that bind to PrPC and stabilize its structure. Here, I investigate the influence of POM1 and POM6, two monoclonal antibodies, on the flexibility and the interaction with the membrane of PrPC by molecular dynamics simulations. Results show that antibody binding limits the range of orientations of the GD with respect to the membrane and decreases the distance between PrPC and membrane. Furthermore, the GD flexibility (Ilie & Caflisch 2022) and the interactions of the flexible tail and the GD are modulated differently by the two antibodies (Ilie et al. 2022). Additionally, I introduce a novel computationally-based approach towards the design of peptides that could modulate the activity of the cellular prion protein.
14:30 - 14:45
Digital Twins with the Multiscale Modelling and Simulation Framework and MUSCLE3 -
Lourens Veen - Netherlands eScience Center visit profile
Presentation category:
Abstract details
Digital Twins of complex systems are an increasingly popular research topic in many branches of computational science. Modelling systems like the human body, an ecosystem or a fusion reactor requires modelling many individual processes taking place in the system, and then combining those models into a coupled simulation. Such a coupled simulation captures the interactions between the processes, allowing complex behaviours to emerge and to be investigated in-silico. The Multiscale Modelling and Simulation Framework (MMSF) describes which kinds of scale-overlapping and scale-separated couplings can exist between processes occurring on different scales in time and space, and therefore how to create couplings between the submodels in a coupled simulation. The Multiscale Coupling Library and Environment 3 (MUSCLE3) software is a user-friendly coupling framework that makes it easy to build coupled simulations out of simulation programs written in Python, Fortran and C++. In this presentation, we will give an overview of the MMSF and MUSCLE3, showcase current uses and the latest developments, and take a peek into the future.
14:45 - 15:00
Advancing Wind Farm Modeling through Fluid Physics and High Performance Computing -
Richard Stevens - University of Twente visit profile
Presentation category:
Abstract details
Wind turbines interact with the environment across a broad range of length scales, spanning from millimeters (viscous scales) and meters (wakes and tip vortices) to "geophysical scales" of hundreds of meters (inter-turbine spacing) up to tens of kilometers (windfarms). This wide range of scales presents significant challenges in theoretical analysis and numerical simulations of windfarm dynamics. Understanding these interactions is crucial to improve windfarm design and operation. Modeling windfarm-atmosphere interactions is an illustrative example of a multiscale physics problem crucial for the renewable energy transition. Successfully addressing this challenge necessitates a synergistic approach involving the development of novel physical models and high-performance computing strategies. The performance of large windfarms depends on the development of turbulent wind turbine wakes and the interaction between these wakes. Turbulence is crucial in transporting kinetic energy from the large-scale geostrophic winds in the atmosphere towards heights where windfarms can harvest this energy. High-fidelity simulations offer insights into the interaction between windfarms and the atmosphere. While early studies centered on 'ideal' scenarios, recent efforts consider terrain and atmospheric stability. This presentation addresses challenges in modeling and computing challenges in understanding the multi-scale modeling challenges of capturing the windfarm-atmosphere interactions.
15:00 - 15:30
Coffee Break
15:30 - 16:00
Keynote: Infectious Disease Data Analytics and Modelling: Mpox, COVID-19
Location: Plenary Room - Juliana Congreszaal
Jacco Wallinga - RIVM, LUMC visit profile
Presentation category: Data-driven methods
Abstract details
An epidemic of a new infectious disease can overburden the health care system and disrupt society. This presents a challenge for policy makers who have to decide on control measures and balance these against the disruptive effects. This presents also a challenge for scientific advisors to policy makers, who have to learn from incoming data and at the same time provide advice and indicate the many uncertainties pertaining to this advice. Modelling can be a way to learn from incoming data, make projections of future scenario’s while indicating the uncertainties. During the COVID-19 pandemic a wide range of different models have been developed and used. In 2022 there was a rapid global spread of mpox that resulted in a different range of models. The difference in models between COVID-19 and mpox is mainly due to the difference in the distribution of the contact rates. We highlight three challenges. A first is how to make models reproducible by external parties that do not have access to data streams from hospitalization and notification records which contain privacy sensitive information. A second is how to make models comparable between countries and regions such that performance of country-specific models can be compared on a much larger scale than before. A third challenge is to maintain and expand the expertise in data analytics and modelling of infectious diseases in the Netherlands.
16:00 - 16:30
Special Lecture: Computational Science and Engineering: Looking back and looking forward
Location: Plenary Room - Juliana Congreszaal
Margot Gerritsen - Stanford University / Women in Data Science Worldwide visit profile
Presentation category: Data-driven methods
Presentation summary
Computational Science and Engineering: Looking back and looking forward (Stanford University) I've had a wonderful (nearly) 40 years as computational mathematician. With the rapid increase in compute power, the development of many new tools and algorithms, and the penetration of computational mathematics in nearly every field, it's been quite the ride. Even though much seems to have changed, much also stayed the same. I look forward to talking with you about the past (and my past), present and a little bit also about the future of this incredible and exciting field we are all in together.
Abstract details
BIO: Margot is Professor [Emerita] at Stanford University and the Executive Director of Women in Data Science Worldwide, a non-profit she spun off from Stanford campus in 2023. She obtained her Ir degree (Technische Wiskunde) at Delft University of Technology in 1990, and then moved to the US in search of sunnier and hillier places. She received her PhD in Scientific Computing & Computational Engineering at Stanford in 1996. Before returning to Stanford in 2001 as faculty member in Energy Science & Engineering, Margot spent 5 years as lecturer at the University of Auckland in New Zealand. From 2010-2018, Margot was the Director of the Institute for Computational & Mathematical Engineering at Stanford,and from 2015 - 2020 the Senior Associate Dean of Educational Affairs in Stanford's School of Earth Sciences. Margot's research interests span a range of topics in computational mathematics, including numerical analysis, Computational Fluid Dynamics, data science, and topics including coastal ocean modeling, reservoir simulation, renewable energy and sustainability. Margot is a Fellow of SIAM (Society of Industrial & Applied Mathematics), and received honorary doctorates from Uppsala University in Sweden and Eindhoven University of Technology in the Netherlands. Margot lives in sunny and hilly Oregon with her husband Paul.
16:30 - 16:45
Closure & Awards
Location: Plenary Room - Juliana Congreszaal
Johan Mentink visit profile
16:45 - 18:00
Drinks