Research projects

Machine Learning for Space Physics

Deep Learning and Gaussian Processes

Can we use solar images and in-situ satellite data to train models that can predict Space Weather events? Can Machine Learning help in discovering new physics? The first book on the topic here.

Solar Wind turbulence

Kinetic physics of turbulent dissipation regime

How does turbulent energy dissipate in a collisionless plasma? What is the role and the interplay between coherent structures, magnetic reconnection, wave-particle interaction at sub-ion scales?

Uncertainty Quantification

Adaptive Sampling, Markov Chain Monte Carlo

What is the best strategy to sample input parameters from an high-dimensional space? How to interpret the result of a deterministic simulation in a probabilistic setting?

Computational Methods for Vlasov Equation

Spectral methods, Particle-in-Cell

How to combine the fluid and kinetic description in an optimal way, with low computational cost?

Machine Learning for Space Physics

Coupling physics-based simulations with Artificial Intelligence

I am leading a team that is pioneering the use of Machine Learning in Space Physics on several fronts.
In our team, we exploit a unique combination of expertise in Space Physics, Computational Physics, Machine Learning and Data Analysis.
We aim at enhancing the current state-of-the-art simulations for Space Weather, by using prior knowledge gathered from historical satellite data. Several Machine Learning techniques will be used for data-mining, classification, and regression. The long-term objective of the project is the creation of a portfolio of data-enhanced reduced models, along with automated rules for model selection.

All our codes and data are publicly available.

Sponsors and Grants

Alt Text Alt Text Alt Text Alt Text Alt Text

NWO is CWI is the Dutch Research Council. INRIA is the French Institute for Research in Computer Science and Automation. AIDA and ESCAPE are Horizon-2020 projects funded by the European Commission. SOARS, Significant Opportunities in Atmospheric Research and Science, is an undergraduate-to-graduate bridge program sponsored by NSF.

More details on our Machine Learning projects are available here.
Back to the top

Uncertainty Quantification

Understanding how uncertainties associated to input parameters propagate through a non-linear simulation and produce a stochastic output form the core of a subject called Uncertainty Quantification. I am interested, in particular, in the so-called non-intrusive approach, where one can use a black-box simulation model and run an ensemble of such simulations to produce a probabilistic output. A major research question is then how to optimally sample the input parameter space, and whether one can hope to do any better than (quasi-) Monte Carlo methods. I have introduced a new adaptive sampling method that consistently beats Monte-Carlo also for very large dimensions.

Relevant publications:

Solar Wind Turbulence

Kinetic physics of turbulent dissipation regime

Models of Solar Wind acceleration and coronal heating predict that the flux of energetic particles moving away from the Sun has to undergo wave-particle interactions in order to locally gain energy and not cool adiabatically, as it is observed. The mechanism behind those interactions is still not clear; in the "turbulent cascade" picture the typical scale of those phenomena would be within the small dissipation regime. That means that any fluid approach would miss important informations about the thermal properties of the particles, and a kinetic approach is needed. I am investigating the physics governing such phenomena by computer simulations (mostly Particle-in-cell).
Relevant publications:

Vlasov Equation

I have a strong expertise in Particle-in-Cell (PIC) methods, which are used by the overwhelming majority of plasma physicist to study the kinetic properties of hot plasmas. One of the PIC code that I have more intensively used is based on the implicit moment method. It has the feature of relaxing some stability constraint, with respect to standard PIC codes.
Unfortunately, PIC suffers from noise and cannot afford an high resolution in phase space. Therefore, for some physical problems (such as involving energy dissipation), one has to use more expensive methods aimed at directing solving Vlasov equation. I am interested in novel computational methods to solve the Vlasov equation.
A promising technique that I am investigating is a Galerkin spectral method employing Hermite functions.

Relevant publications: