Student Posters Repository

Radiation hydrodynamics (RHD) is an essential component in numerous scientific simulations like astro-, particle-, and laserphysics and up to this day a challenging field to model with simulations because of its physical complexity as well as its numerical challenges. Particular fields of interest, in which the handling of anisotropic radiation is required are namely star formation, planet formation and supernova research. Finding a scheme that combines the necessary physical complexity while remaining numerically affordable is therefore imperative for simulations in these fields.

Moment methods are popular for radiation hydrodynamics simulations due to their reduced computational cost but have to be used with care, because of the approximations applied in their derivation. This project is about the implementation of a moment method for radiation transport, namely the M1 method into the widely used astrophysical code PLUTO. The crucial advantage of this method over the previously implemented radiation transfer model is its capability to handle anisotropic fields of radiation.
Yet, the additional complexity of the model results in higher demands for computational resources. So far, the scheme is fully parallelized using MPI and the PETSc library  and  tested  in  up  to  three  spatial  dimensions.  First  simulations display the physical advantages of the M1 method and make it a promising tool for further studies. Ongoing work lies in large scale performance evaluation and optimization to increase its usability for scientific studies.

In nuclear fusion devices, such as tokamaks and stellarators, plasma properties (density/temperature) at the plasma edge and in particular the scrape-off layer (SOL) significantly influence magnetic confinement within the plasma core. Steep gradients and large amplitudes of perturbation are present in the plasma edge and the so-called scrape-off layer region. Hence, understanding the tokamak edge is crucial for achieving successful magnetic confinement. The newly developed "PICLS" code specifically addresses SOL physics. Numerically, it is based on a Particle-in-Cell (PIC) Monte Carlo algorithm with hybrid OpenMP/MPI parallelization. Physically, a gyrokinetic (GK) model is implemented that allows to de-couple fast gyro-angle movement and thus enables significant speed-up compared to fully kinetic models. However, the validity of the GK approach within the SOL still needs to be investigated further. Up till now we implemented a 1D1V model that could successfully be tested for various plasma wall sheath studies. As a next step, the model will be extended towards three spatial dimensions to tackle more realistic problems.

Ordinary differential equation (ODE) models have become an important tool in systems biology. Mechanistic ODE models allow for the integration and mechanistic interpretation of heterogeneous data. Understanding complex diseases and interpreting multi-omics datasets requires the use of large-scale models with thousands of unknown parameters, the estimation of which is computationally very costly.

We are developing and applying tools for scalable parametrization of large-scale ODE models. Our AMICI toolbox [1] allows for scalable simulation and sensitivity analysis of ODE models. For parameter estimation, we developed a C++/MPI library [2] for scalable parallel model simulation and gradient-based optimization.

The integration of these tools allows us to parametrize models with thousands of states and parameters, a scale which has so far been infeasible in systems biology [3,4]. Such large-scale models could have a huge impact in systems biomedicine by providing in silico predictions to be used for drug target identification, for personalized medicine or even virtual clinical trials.

References:

[1] Advanced Multilanguage Interface to CVODES and IDAS (AMICI), https://github.com/ICB-DCM/AMICI/
[2] Parameter estimation library parPE, https://github.com/ICB-DCM/parPE
[3] Schmiester et al., Efficient parameterization of large-scale dynamic models based on relative measurements, 2019, bioRxiv, 579045.
[4] Fröhlich et al, Efficient Parameter Estimation Enables the Prediction of Drug Response Using a Mechanistic Pan-Cancer Pathway Model, 2018, Cell Systems, 7(6):567-579.e6.

More than 80% of the stars in the Universe form in very crowded environments called star clusters. Such systems may contain a number of stars which ranges from few hundreds to few millions, and their size is of the order of 1pc (3.086E13 m). In such dense and crowded environments, dynamical interactions and collisions between stars may trigger a plethora of interesting events, many of which are still being studied and matter of debate. In my research, I simulate such systems using parallel computing, and in particular I exploit GPUs, which are perfectly suited to simulate the gravitational interactions of such systems since they can efficiently parallelize all the Newtonian forces computations that each particle exerts on each other particle. Many complex algorithms and numerical tricks such as regularization and block-timestep are needed to perform such simulations. Furthermore, I need to simultaneously simulate all the needed astrophysical processes such as single/binary stellar evolution, stellar winds, star collisions, etc. In my last project, I try to understand how dynamics of star clusters may lead to the formation of gravitational wave sources (e.g. binary black holes) similar to the recently detected ones and see how it influences the properties of such systems.

Molecular dynamics (MD) simulations are employed to investigate the ability of atmospheric particles consisting of a plethora of organic compounds and water to form glassy phases and initiate ice nucleation in low temperatures. Supercooling dynamics was employed and managed to probe an observed behaviour of organics that can enhance or inhibit Ice Nucleation. We show that organic components do not partition equally to the ice and aqueous phases. Instead, organic-rich particles preferentially remain unfrozen and have the potential to influence aerosol-cold cloud interactions and climate.

Communication overhead in parallel systems can be a significant bottleneck in scaling up parallel computation. In this paper, we propose event-triggered communication methods to reduce such communication overhead for numerical simulation of partial differential equations. As opposed to traditional communication in which processing elements exchange data at every iteration of the numerical algorithm, the main idea behind event-triggered communication is to exchange data only when necessary as dictated by a suitably defined event. We show through numerical experiments that these methods have potential to reduce simulation time. Strong scaling plots show that the approach may be effective on large machines as well.

The computational modelling of chemistry from first principles in general is hard: the Schrödinger equation must be solved for molecular systems, and such a task is prohibitively expensive due to the exponential scaling of the simulation cost with respect to number of particles. The Born-Oppenheimer approximation is used to facilitate such calculations, separating nuclear and electronic degrees of freedom in what is called an adiabatic propagation. Yet, on systems where there might be several electronic states coming into play, such as metallic surface systems, the approximation break and nonadiabatic dynamics come into play, rendering the calculations unfeasible within reasonable computation times. The aim of my project was to develop a novel method for such simulations, which is able to properly model all the interesting physical phenomena of such reactions while avoiding the large computational cost.

The “Reconstructed Image from Simulations Ensemble” (RISE), based on Monte Carlo simulation techniques, statistical physics concepts and parametric modelling
of the imaged target, provides a novel reconstruction method for emission tomography. The method simulating a huge ensemble of possible image reconstructions and
choosing the “optimum” solution based on a goodness criterion, provides reconstructions of superior image quality as compared to the conventional methods. Although
RISE has proven well suited in cases of noisy data, it could not be implemented without utilizing High Performance Computing resources. The capabilities of GPUs to provide
high performance computation in such simulation problems enabled the implementation of the method in emission tomography for image reconstruction of real samples.