I am a development lead for Phantom, a high-performance astrophysical SPH code for simulating a wide array of astrophysical problems.

Simulations are computationally expensive. Architectures and super-computing clusters continue to increase in scale. Making efficient use of these resources requires highly-effective parallel schemes, so that memory and communication overhead do not bottleneck a computation. I am interested in developing shared-memory threaded schemes that can scale to large core numbers, and in developing distributed computing algorithms for SPH codes that can scale to hundreds or thousands of nodes of a cluster.

All simulations are underpinned by the quality of their methods. I conduct experimental studies to understand and proof the numerical properties of SPH, to identify items such as the root cause of numerical errors and in what physical regimes a numerical method may or may not be suitable. For example, I have studied the convergence properties of SPH on the linear and non-linear evolution of the Kelvin-Helmholtz instability, demonstrating the numerical conditions under which convergent solutions may be obtained.

Astrophysics is a soup combining a multiple types of physics. My research involves creating and testing multi-physics SPH algorithms, that is, SPH methods that simultaneously solve hydrodynamics which is coupled to other physical processes. A key line of research in this area has been the development of magnetohydrodynamics methods for SPH.

Sarracen is our Python library for SPH visualization and analysis. Our goal is to provide support for visualization and analysis of SPH data in a SPH-specific way using the smoothing kernel. Sarracen can be used interactively in a Jupyter notebook environment, and supports multi-threaded and GPU (through CUDA) visualization.

Banksformer is our transformer-based machine learning model for the generation of synthetic financial transaction sequences. Financial transaction sequences are a complex series of multivariate, multi-modal temporal events that occur at irregular intervals. Banksformer is designed for general banking data, such as personal bank accounts. It supports multiple classes of transactions, and contains specific optimizations to robustly capture time and date patterns.

Our agent-based approach to generate financial transaction sequences simulates the behaviour of individuals as it pertains to create a realistic history of financial transactions.

We use generative deep-learning techniques, such as generative adversarial networks (GANs), for the creation of a variety of data. DG-GAN is our model for generating demographic data.