Summary
Plasma physics has seen a long tradition of deriving simplified models such as magnetohydrodynamics through a process known as closure: starting from computationally demanding kinetics and following various analytical approximate schemes resulting in fluid-type models which are less accurate but more tractable. Due to smaller computational footprint these models have found applications in space weather modelling and fusion. The problem is that in many interesting applications, e.g. where collisions between particles are rare, the analytic closures have limitations.
The goal of STRIDE (Systematic Techniques for Robust Inference and Data-driven Explainable closures for plasma) is to use machine learning to construct models with fewer degrees of freedom that describe kinetic processes relevant for Geospace Environmental Modelling (GEM), such as magnetic reconnection. Corrections to fluid-type models will be learned with deep neural networks and equation discovery and tested in numerical simulations. The important challenge involves understanding how such surrogates can be made robust against out-of-distribution shifts (i.e. different physical conditions) and numerical instabilities. Thus the closures will be first trained on data generated by high fidelity physics-based model, e.g. kinetic Particle-in-Cell simulations, for a specific set of parameters and then transfer learning will be applied to a different set of parameters to improve robustness. Uncertainty quantification and the ability to generate extremes will be investigated.
The scientific question to be addressed here include: how can we interpret the trained models using explainable AI, what are the optimal ways of performing transfer learning and fine-tuning on the observational data, are there physical considerations (such as conservation laws or symmetries) that can reduce the costs associated with training such machine learning models? Trained models will be made open-source to foster research.
The goal of STRIDE (Systematic Techniques for Robust Inference and Data-driven Explainable closures for plasma) is to use machine learning to construct models with fewer degrees of freedom that describe kinetic processes relevant for Geospace Environmental Modelling (GEM), such as magnetic reconnection. Corrections to fluid-type models will be learned with deep neural networks and equation discovery and tested in numerical simulations. The important challenge involves understanding how such surrogates can be made robust against out-of-distribution shifts (i.e. different physical conditions) and numerical instabilities. Thus the closures will be first trained on data generated by high fidelity physics-based model, e.g. kinetic Particle-in-Cell simulations, for a specific set of parameters and then transfer learning will be applied to a different set of parameters to improve robustness. Uncertainty quantification and the ability to generate extremes will be investigated.
The scientific question to be addressed here include: how can we interpret the trained models using explainable AI, what are the optimal ways of performing transfer learning and fine-tuning on the observational data, are there physical considerations (such as conservation laws or symmetries) that can reduce the costs associated with training such machine learning models? Trained models will be made open-source to foster research.
Unfold all
/
Fold all
More information & hyperlinks
| Web resources: | https://cordis.europa.eu/project/id/101148539 |
| Start date: | 01-09-2024 |
| End date: | 31-08-2026 |
| Total budget - Public funding: | - 175 920,00 Euro |
Cordis data
Original description
Plasma physics has seen a long tradition of deriving simplified models such as magnetohydrodynamics through a process known as closure: starting from computationally demanding kinetics and following various analytical approximate schemes resulting in fluid-type models which are less accurate but more tractable. Due to smaller computational footprint these models have found applications in space weather modelling and fusion. The problem is that in many interesting applications, e.g. where collisions between particles are rare, the analytic closures have limitations.The goal of STRIDE (Systematic Techniques for Robust Inference and Data-driven Explainable closures for plasma) is to use machine learning to construct models with fewer degrees of freedom that describe kinetic processes relevant for Geospace Environmental Modelling (GEM), such as magnetic reconnection. Corrections to fluid-type models will be learned with deep neural networks and equation discovery and tested in numerical simulations. The important challenge involves understanding how such surrogates can be made robust against out-of-distribution shifts (i.e. different physical conditions) and numerical instabilities. Thus the closures will be first trained on data generated by high fidelity physics-based model, e.g. kinetic Particle-in-Cell simulations, for a specific set of parameters and then transfer learning will be applied to a different set of parameters to improve robustness. Uncertainty quantification and the ability to generate extremes will be investigated.
The scientific question to be addressed here include: how can we interpret the trained models using explainable AI, what are the optimal ways of performing transfer learning and fine-tuning on the observational data, are there physical considerations (such as conservation laws or symmetries) that can reduce the costs associated with training such machine learning models? Trained models will be made open-source to foster research.
Status
SIGNEDCall topic
HORIZON-MSCA-2023-PF-01-01Update Date
25-11-2024
Images
No images available.
Geographical location(s)
