Student Posters Repository

The formation of Population III (Pop III) stars is a critical step in the evolution of the early universe. To understand how these stars affected their metal-enriched descendants, the detail sof how, why and where Pop III formation takes place needs to be determined. One of the processes that is assumed to greatly affect the formation of Pop III stars is the presence of a Lyman-Werner (LW) radiation background, that destroys H₂, a necessary coolant in thecreation of Pop III stars. Self-shielding can alleviate the effect the LW background has on the H₂ within halos. In this work, we perform a cosmological simulation to study the birthplaces of Pop III stars, using the adaptive mesh refinement code Enzo. We investigate the distribution of host halo masses and its relationship to the LW background intensity. Compared to previous work, halos form Pop III stars at much lower masses, up to a factor of a few, due to the inclusion of H₂ self-shielding. We see no relationship between the LW intensity and host halo mass. Most halos form multiple Pop III stars, with a median number of four, up to a maximum of 16, at the instance of Pop III formation. Our results suggest that Pop III star formation may be less affected by LW radiation feedback than previously thought and that Pop III multiple systems are common.

Operation batching is an important way to accelerating the execution of neural networks having dynamic computation graphs. There have been some strategies proposed to batch the operations automatically and execute the batched operations. But there are still opportunities for further optimization in computing performance. In my research, I found there are least 3 further optimizations: 1) The calculation of the parameters’ gradients could be executed in a more efficient way; 2) Overhead coming from data copy could be eliminated if the memory stores the arguments of a batched nodes could be allocated in continuous space in advance; 3) Nodes that don’t depend on each other in the computation graph generated by the automatic operation batching strategies could be executed simultaneously in a task-parallel way. Optimization 1) has already been tried and evaluated on 5 different NLP task benchmarks. Improvements in computing performance have been observed in some cases while decrement also happened in other cases. Profiling has been done and the result shows that Optimization 2) and 3) are promising.

We present a theoretical and numerical study on the stability of the interface between two fluids in a Hele-Shaw cell. Specifically, we consider the effect of a geometric taper in the direction of flow, across a range of capillary numbers $Ca$. We supplement linear stability results with fully-resolved 3D simulations (thus computing the flow field in the Hele-Shaw cell) carried out using the InterFoam solver in OpenFOAM, which employs the volume-of-fluid method to evolve the fluid-phase-field and enhances accuracy via an interface-compressing term in the continuity equation.

Three types of Hele-Shaw cells are considered: diverging, converging and parallel. Al-Housseiny et al.\ found a critical $Ca$ in the converging cell, below which the interface is stabilized. We extend this analysis by introducing a local $Ca$, which varies along the flow direction in tapered cells. Based on the difference between the critical $Ca$ and the inlet or the outlet $Ca$, the (in)stability scenarios are divided into three regimes for each cell. Results from our 3D simulations show good agreement with the theoretically predicted (in)stability regimes and linear growth rates well with this theoretical analysis, validating our classification. 

My research interest focuses on parallel 3-D mesh refinement for finite element electromagnetics (FEM) with tetrahedron on CPU and GPU, using tessellation shaders. Currently, 3-D models require vast numbers of elements. Thus, the benefits of this new approach are for developing high-performance parallel mesh refinement algorithms, efficient and accurate solutions for three-dimensional (3-D) modern applications. The main objective is to purpose is to propose a 3-D parallel mesh refinement model suitable for FEM electromagnetics with tetrahedron to be executed by a parallel engine, that is, investigate algorithm development for systems of heterogeneous multiprocessors; and detailed communication cost analysis of such methods.

In modern dynamical system modelling, finding coordinate transformation for representing highly non-linear dynamics in terms of approximate linear dynamics has been of crucial importance for enabling non-linear control, estimation, and prediction. Recently developed interest in Koopman operator theory has shown that its eigenfunctions can provide such coordinates that intrinsically linearize the global dynamics But finding and representation of such eigenfunctions have been challenging. The present work leverages deep learning methods, specifically Recurrent Neural Networks (RNNs) for discovering the Koopman eigenfunction representations and exploit RNNs ability to model temporal dependencies, to allow multi-step evolution of the dynamics, as long forecasting for such systems still remains a major challenge. Current work is an incremental work on the network architecture, which is interpretable in terms of Koopman theory and parsimonious, allowing augmentation to the lacking interpretability to deep learning architectures, while capturing the fewest meaningful eigenfunctions. Some other challenges related to modelling such architectures are discussed in future work.

Snow cover is a key component of the global climate system, is crucial for the water supply of large parts of the world's population, and has economic importance e.g. for hydropower generation or in the winter tourism industry. We simulate the temporal and spatial evolution of the mountain snow and ice cover on regional scales, which requires solving for the energy and mass fluxes between the atmosphere, the snow cover and the ground. Accounting for small-scale topography-driven processes such as radiation fluxes from surrounding terrain or wind-driven or gravitational snow redistribution requires high resolution simulations, associated with large computational demands especially when performing operational ensemble forecasts or long-term climate change scenario simulations. As the current model version is only weakly parallelized, currently a reimplementation of the model is being initiated which should exploit state-of-the-art HPC methods.

Nowadays, billions of people use a huge variety of online services on a daily basis. These services are successful because they provide personalized results depending on the user profile (user’s identity associated with their personal data).  Majority of online services relies on third party cloud services either for data storage or for data processing. A leakage of user profile, which contains sensitive information, may jeopardize the end user privacy.

The Data partitioning approach could be considered to overcome the issue. This approach consists of dividing a sensitive user profile into non-sensitive components. Then the code is partitioned based on data partitions and it is deployed either on multiple server or on multiple enclaves within a same server. However, this approach is challenging to implement, because the developer manually partitions the code and tracks the data flow. Even a small careless mistake may lead to unintended data leakage.

My PhD thesis has the goal of developing a domain specific language (DSL) to enforce privacy. The language consists of annotating C code to express data sensitivity. Furthermore,  the DSL ensures correctness using static analysis, automatically partitions the code and deploys it. We modify the LLVM compiler for the implementation. Our language reduces coding time and attack surface of the code, prevents unintended data leakage and ensures portability for different frameworks.

        Over the past two decades, freshwater including lakes, reservoirs, streams and rivers are receiving accumulating attention as important global methane sources, especially for regions as sensitive to global warming as the Arctic. Nevertheless, studies have shown large uncertainties in the estimation of freshwater methane emissions. Our research is to simulate lake methane emissions using a one-dimensional process-based model, called the Arctic Lake Biogeochemistry Model (ALBM). It mainly consists of several modules: the water/sediment thermal module, the water/sediment biochemical module, the gas diffusion module and the methane ebullition module. We currently have got the Finnish lake database (SYKE) including the basic physical information of 214,995 lakes, and observation of methane fluxes from 177 lakes from 1998 to 1999. There are two steps of modeling: 1) parameter calibration using observations for individual lakes using a Monte Carlo calibration method; 2) regional simulation of 214,995 lakes using SYKE data product for both historical and future results. Since the model is site-level, the study requires relatively extensive computing resources to run the simulations, which would benefit from more efficient high parallel computing approaches.