Student Posters Repository

Program synthesis has attained significant attention as a powerful tool to enhance software development via automation. However, the most demanding obstacle in this domain is the efficiency of search algorithms that restrict the practical application of this approach.

Refinement-Typed Genetic Programming (RTGP) has been proposed as a promising approach in this field, yet it has not been thoroughly implemented and evaluated. To address this gap we focus on the practical implementation and comprehensive evaluation of RTGP. RTGP merges Liquid Types with Genetic Programming. This work aims to prune and optimize the search space of solutions, while also increasing the expressiveness available to programmers. It allows them to concentrate on defining the most suitable specifications for what the program should accomplish or how it should behave. This focus significantly reduces development time and effort. Consequently, it facilitates the production of high-quality software without requiring programmers to direct their attention and time towards the means of problem- solving. 
The computational demands of RTGP, particularly in exploring search spaces, align perfectly with high-performance computing (HPC) environments. Leveraging HPC’s parallel processing capabilities and extensive computational resources could potentially improve the efficiency and feasibility of RTGP, making RTGP a practical tool for real-world applications.

Our research aims to design a programming language that implements this approach, evaluate the performance and expressive power using a program synthesis benchmark, and compare it to other existing program synthesis techniques

We propose a deep generative model for 3D shapes that incorporates structural mechanics parameters,

and a dataset of 10114 shapes created by topology optimization. Our model is based on DeepSDF,

a decoder-type neural network that implicitly represents shapes as signed distance functions (SDFs). We

extend DeepSDF to condition the shape generation on structural mechanics parameters, such as absorption energy. We also introduce positional encoding to improve the spatial

resolution of the model. Our dataset consists of various 3D shapes computed by a linear topology optimization

method using the Building-Cube method. We use the impact absorption energy as a quantitative indicator of the

structural performance of the shapes. We train our model on the dataset and evaluate its ability to generate

3D shapes reflecting structural mechanics parameters. Our results indicate that our model can produce 3D

shapes with high fidelity and diversity, and achieve an average reconstruction accuracy of 96.5% for the test

shapes. Our model and dataset open up new possibilities for 3D shape generation and structural design using

deep learning.

Droplets are ubiquitous both in daily life and industrial applications from self-cleaning surfaces, to ink-jet printing. Due to their multi-scale nature, with fluid dynamics describing the behaviour of the bulk and molecular interactions dictating the dynamics at the triple phase point (the contact line), predicting the dynamics of droplets is a challenging problem. Numerical solutions to arising Partial Differential Equations are hard to obtain, since the equations are stiff. In this work a matched asymptotics approach is developed, simplifying the problem to a reduced order model of coupled differential equations which is naturally solved using spectral methods. Comparison with numerical solutions to the full governing equations show that the reduced order model provides highly accurate solutions at a fraction of the computational cost. In turn, this allows the study of evaporative effects on the droplet dynamics and paves the way for data-intensive tasks - an example being solving the inverse problem, which arises in industries where fine control, via optimal choice of the substrate heterogeneity profile, of small fluid substances is required.

The coronary arteries surround the heart and supply blood to the muscles that pump oxygen and nutrients to the heart. Anomalous Aortic Origin of the Coronary Artery (AAOCA) is a congenital anomaly, the second leading cause of sudden death in young athletes. The hemodynamic analysis is necessary due to the inadequacy of clinical imaging techniques in determining the problems caused by anomalies in blood flow.

The left coronary artery (LCA) geometry is reconstructed from CT images of a healthy 48-year-old male using 3D Slicer®. Two AAOCA models with acute take-off angles of 38° and 23°, and elliptical ostia are developed. Computational Fluid Dynamics (CFD) simulations are employed, utilizing both Newtonian and non-Newtonian rheological models. The OpenFOAM^® ESI PIMPLE solver is used to simulate transient, incompressible, and viscous blood flow phenomena, including systole and diastole phases over two cardiac cycles. Simulations are run in parallel on the HPC system at the National Centre for High Performance Computing (UHeM) at ITU. Hemodynamic indexes are utilized to investigate how the take-off angle and ostium shape of the LCA affect blood flow in individuals with AAOCA. These indexes include velocity streamlines, wall shear stress (WSS), time-averaged wall shear stress (TAWSS), transverse wall shear stress (transWSS), and fractional flow reserve (FFR).

The Eulerian method is a numerical method using a spatially fixed mesh and is suitable for large-scale simulation, such as structural simulation that requires a huge number of continuum elements. However, no method for solid contact analysis with slip has been proposed in conventional Eulerian methods. In this study, we propose an Eulerian solid contact analysis method that takes slippage into account for a situation in which a structure and a rigid jig come into contact with each other. Furthermore, we apply the proposed method to the elastoplastic analysis of a crush box and verify the usefulness of the proposed method.

We present a preliminary study of timescales of many-body fast neutrino flavor conversions in core-collapse supernova, paying particular attention to distributions of Loschmidt echo crossing times (intimately connected to dynamical phase transitions in non-equilibrium systems) determined by time evolution of the many-body Hamiltonian. Starting from a tensor-product state of systems with N neutrinos (N/2 electron-type and N/2 heavy-type) the Loschmidt echo crossing times, tL× , are found to exhibit two distinct time scales that are exponentially separated. Distributions are found to become increasingly well described by the sum of two stable distributions.

The paramount role of enzymes in an enormous amount of biological processes is undisputed. However, the origin of their immense catalytic power remains a hotly debated topic. Among many of the proposed explanations, two are in the forefront – one claims that the driving force of enzyme catalysis are dynamical effects and the other that enzymes work through preorganized electrostatics. With our studies we aim to shed some additional light on this important topic by devising a simple, yet efficient multiscale computational model, which allowed us to quantify the effect of electrostatics on (monoamine oxidase) enzyme catalysis. In addition, we also looked at our systems from a different angle, by representing the reacting moiety with a dipole moment vector and the solvated enzymatic environment with an electric field vector. In doing so, we can calculate the established free energy of the interaction between the two vectors and look at the contributions of individual amino acid residues.

With this approach we investigated several different reactions catalysed by monoamine oxidase (MAO) enzymes. We successfully elucidated the importance of electrostatic interactions in the MAO-A catalysed degradation of phenylethylamine, compared the effects of electrostatics between the two MAO isoenzymes and performed an in-depth study of Brunner syndrome, a debilitating neurological disorder caused by a mutation in the MAO-A encoding gene.

With the developed multiscale computational model, we confirmed the decisive role of electrostatic interactions in the catalytic function of MAOs, and we are working on expanding our research to a wider variety of enzymes. The results produced so far are an important contribution towards resolving one of the most important unanswered questions in biochemistry, namely elucidating the driving force behind enzyme catalysis.

This study combines the background of numerical analysis for Discontinuous Galerkin Methods for solving hyperbolic-parabolic PDEs and numerical algorithms for the development of a high-order CFD code. The main objective is to derive a scheme that respects the non-linearity of the compressible Navier-Stokes problem for up to a desired polynomial order of approximation. The targeted problems at the moment mainly concern subsonic external aerodynamics with solid wall bodies, i.e. bluff bodies, airfoils etc. The implementation of the algorithm uses special abscissa for interpolation of the flow which minimizes the Lebesque constant and allows to reach very high polynomial orders. The code is written in modern Fortran and follows a pure MPI path to parallelize the heavy computational load.  Upcoming study will involve efficient time integration free of a CFL constraint and acceleration of the matrix multiplications which are the major computing intensive part.

In the design of a fusion power plant, particular attention must be paid to turbulent transport at the boundaries of the confined region, i.e. the plasma edge and the scrape-off layer (SOL), as this is critical for managing heat and particle exhaust while optimising core confinement. In order to extrapolate current observations to future devices, it is necessary to gain sufficient physics understanding through the use of first-principle codes. The gyrokinetic code GENE-X has been specifically designed to study edge and SOL plasma turbulence in realistic geometries. It is a high-performance code with a hybrid MPI+OpenMP parallelisation and is currently being used on up to 25,000 CPU cores (ongoing GPU porting). In addition to the main plasma, molecules and neutral atoms (neutrals) are present, especially in the SOL, and interact with the plasma through a complex set of collision processes. This has several crucial and non-negligible effects on the plasma parameters and confinement. In order to improve the predictive capabilities of GENE-X, a neutrals model and the plasma-neutrals interactions have been implemented and verified using the method of exact solution. Initial simulations of isolated plasma filaments are being performed, but in order to unleash the reactor-relevant geometries capabilities of GENE-X, the neutrals model needs to be fully parallelised and optimised. In particular, the interplay between the parallelisation of the five-dimensional plasma model and the three-dimensional neutrals model is of critical importance in order to achieve optimal load balancing.

In recent years there has been an explosion in quantum algorithm development for Noisy Intermediate-Scale Quantum (NISQ) devices. The dominant approach has been in the form of Variational Quantum Algorithms (VQAs) which use a hybrid loop of parameterised quantum circuits which are then optimised classically to minimise a given cost function. A major roadblock in the success of these algorithms is the barren plateau phenomenon which occurs when a model has an exponentially flat parameter space, making the parameter optimisation step infeasible. Recent work has successfully characterised the general conditions that lead to barren plateaus. Examining these conditions has led to the conclusion that if a model is feasible to train (no barren plateau) then the entire optimisation procedure can be classically simulated efficiently. This work seeks to realise a proof of concept for this theoretical conclusion by developing a module that allows for the classical pre-training of the quantum parameters for the Quantum Approximate Optimization Algorithm (QAOA), one of the most promising VQA algorithms, on the MaxCut problem. The model leverages known structures in the hamming graph for the solution strings, allowing the cost function to be approximately computed in sub exponential time. While initial testing using serial python code has proved successful, the computation is still demanding and future work aims to develop an efficient OpenMP parallelised code which will allow further testing.