About me
I'm an applied mathematician and engineer working on data-driven simulation of physical systems. I combine reduced-order modelling, data assimilation, and machine learning with physics-based solvers across diverse engineering and scientific applications. Ultimately, I want to develop technologies that help us understand and address the challenges of climate change and contribute to a more sustainable future for everyone.
What I do
Simulation Systems
I build cheap alternatives to the traditional, expensive simulation pipelines evaluating real-world physical systems, such as weather forecasting. Instead of requiring hours' (or days') worth of computations, reduced-order models take minutes to run, while still capturing the essential physics of the problem.
Forecasting Pipelines
These techniques combine sparse, noisy observations with a physical model to produce unknown quantities of interest and their uncertainty. A typical example: given a few imperfect measurements of wind speed and humidity in a given region, what is the most likely state of the atmosphere, and how confident can we be in that estimate? To answer this, I develop Bayesian methods, variational approaches and other staples of modern data science.
Scientific Software
I develop performant scientific software that is designed to deliver meaningful results across a range of engineering and scientific applications. As I tackle primarily large-scale physical problems, HPC capabilities are a central feature of the code I write. I build with reproducibility and documentation from the outset, and I make my work open source so that it can be used, extended, and built upon by anyone interested.
Machine Learning
The massive availability of data has opened up new opportunities for machine learning techniques. These present valid alternatives to traditional physics-based approaches, as they enable the discovery of patterns and relationships even in highly nonlinear regimes. Although it can sometimes be difficult to keep up with a field that seems to develop almost daily, I try to stay on top of these developments and actively explore and implement new methods in my research. As there is still a lot to discover, this is an exciting area to work in: there is still plenty of room to figure out what really works, when, and why.
Current work
Right now that's DigiOcean4Solar, an NWO Vidi project that's building a digital framework to simulate how waves, wind, and floating solar platforms interact at the scale of a whole farm. One island's wake changes the conditions the next island downstream has to be designed for, and modelling the whole farm as a system is what makes that visible ahead of time. I work on the forecasting and uncertainty quantification side of the project, while also developing reduced-order models to make these incredibly expensive simulations computationally feasible. By bringing all these pieces together, we aspire to create an actual digital ocean for solar farms: a virtual farm whose state is known in near real time, continuously updated using information from remote sensors such as satellites.
Selected Work
GridapROMs.jl
Repository of reduced-order models written in Julia. It tackles a wide range of applications, from electromagnetics to nonlinear fluid dynamics, from problems defined on moving geometries to ones partitioned across multiple processors in HPC-friendly environments. Speedups of several thousand times over full-order simulations are possible with these algorithms.
Opal.jl
Data assimilation framework written in Julia. It tackles a wide range of problems, from ordinary differential equations to high-dimensional partial differential equations, from full-order models to reduced-order surrogates. Kalman, particle, and variational filters sit behind one interface, with add-ons for localisation, inflation, and bias correction, and direct integration with ROM pipelines for efficient simulation.
Working Style
I care about turning mathematically sound ideas into scientific software that is reliable, scalable, and useful beyond the problem it was originally designed to solve
Highlights
Developing digital-twin and large-scale multiphysics frameworks for offshore floating solar farms, with a forecasting and uncertainty quantification layer that feeds into how a farm's layout and structures get designed.
Built reduced-order methods, novel tensor-based compression workflows, unfitted finite element techniques, and scientific software for parameterised PDEs, especially in the context of computational fluid dynamics.
Worked on topology optimisation for compliant aerospace mechanisms under manufacturing and stress constraints, connecting numerical methods to engineering requirements.
Thoughts
A mathematical method is only the beginning. I care about turning it into software that is clear enough to understand, robust enough to reuse, and flexible enough to apply beyond the problem that motivated it. Good architecture, testing, examples, and performance are part of the research too.
In GridapROMs.jl, a simple change in the order of for loops turned out to matter more than parallelising them. Looping over mesh cells on the outside and handling every parameter value together on the inside, through a lazy evaluation structure built on Gridap's lazy-array machinery, reuses local finite-element caches across parameter values and cuts memory use dramatically without changing the underlying mathematics. A useful reminder that performance often starts with looking carefully at how a computation is organised.
Reservoir computing is a good example. Only the linear readout of an echo state network needs training, so it skips the expensive automatic differentiation most machine learning methods rely on and stays cheap to retrain. Echo state networks are also naturally suited to capturing time-evolving processes, which makes them a good fit for bias correction, where the model has to keep adapting as new observations arrive.
Machine learning algorithms used as reduced-order modelling tools are developing incredibly quickly, and it can be difficult to keep up. I try to stay close to these developments and explore them in my research, but I don't think we are near having a method that we understand well enough to trust reliably for every problem. There is still plenty to learn about where each approach holds up, where it doesn't, and why.