CD Technical Meeting (ML16): Surrogate models for materials in fusion power plants

Wednesday 9 Sept 2026, 11:00 → 11:30 Europe/London

Eszter Szekely (UKAEA)

Machine Learning, Uncertainty Quantification and Data Science 

 

To accelerate dynamical simulations of irradiated metals, we combine coarse-graining techniques with data-driven surrogates, specifically under conditions that will occur in working fusion power plants which are expensive to test experimentally. Simulating longer timescales will bring us closer to the scales of the lifecycle of fusion power plants. We use autoregressive methods fitted to data gained from coarse-graining molecular dynamics (MD) simulations.  

 

1. Fusion-relevant conditions 

Future fusion power plants could be a safe and low-carbon energy source that would allow humanity to mitigate climate change. Achieving fusion on Earth requires temperatures exceeding those of the Sun’s core, creating extreme thermal conditions. Additionally, the wall materials of the power plant will need to withstand irradiation by high-energy neutrons causing structural and compositional changes over time. Tungsten is a prime candidate wall material due to its high melting point, low erosion rate and low tritium fuel retention. However, its behaviour under prolonged neutron irradiation and high heat flux is not fully understood yet. As the fusion conditions are hard to reach experimentally, computational studies are valuable for the design choices of the power plant. 

 

State-of-the-art MD simulations provide good descriptions of irradiation effects in metals [2]. In this project, we create surrogates for these simulations to reduce computational costs greatly and to make it feasible to simulate longer timescales. 

 

2. Forecasting coarse-grained atomistic descriptors 

Atomistic descriptors (D) are geometric representations of atomic environments. They have been successfully applied as basis functions for developing chemical models [1,3]. Here, we reduce the dimensions of D by coarse-graining descriptors ( ˜D) in a space-averaged way following Ref. [4], but introducing voxelisation. ˜D have been shown to be a good latent space to propagate MD using vector autoregression (VAR) models, and also to fit models predicting global physical properties of interest [ 4]. Thus, the surrogates forecast simulation states with negligible computational costs, and can predict various global observables of engineering interest [4]. However, Ref. [4]’s studies were limited to forecasting MD under fixed physical conditions.  

 

Here, we study the applicability of the method for building models forecasting dynamical simulations of irradiation, transferable to different physical conditions. We build our datasets combining techniques suggested in Ref. [2] and [4]. As our target dynamical systems are more complex, we improve the methodology for accurate forecasting. First, we apply voxelised descriptors which gives access to semi-local observables along with the global ones, and makes it possible to fine-tune the level of coarse-graining in the surrogates. Further, we study how applying more advanced ML models and using different timescales for the fits affect the accuracy of the predictions. We validate our models comparing the predicted observables to physical properties computed from MD simulations. 

 

Acknowledgments 

This work has been funded by the Fusion Futures Programme. As announced by the UK Government in October 2023, Fusion Futures aims to provide holistic sup- 

port for the development of the fusion sector. This work used the ARCHER2 UK National Supercomputing Service (https://www.archer2.ac.uk) and the Pitagora Supercomputer hosted by CINECA under the TADFRM project. 

 

References 

[1] Albert P. Bartók. Gaussian Approximation Potential: an interatomic potential derived from first principles Quantum Mechanics. Ph.D. thesis, University of Cambridge, Engineering Dept, Trumpington St, Cambridge CB2 1PZ, 2010. 

[2] Max Boleininger, Daniel R. Mason, Andrea E. Sand, and Sergei L. Dudarev. Microstructure of a heavily irradiated metal exposed to a spectrum of atomic recoils. 

Scientific Reports, 13(1):1684, Jan 2023. 

[3] Thuong T. Nguyen, Eszter Székely, Giulio Imbalzano, Jörg Behler, Gábor Csányi, Michele Ceriotti, Andreas W. Götz, and Francesco Paesani. Comparison of permutationally invariant polynomials, neural networks, and Gaussian approximation potentials in representing water interactions through many-body expansions. J. Chem. Phys., 148(24):241725, 2018. 

[4] Thomas D Swinburne. Coarse-graining and forecasting atomic material simulations with descriptors. Physical Review Letters, 131(23):236101, 2023. 

 

11:00 → 11:20 Talk: Calibrated Physics-Informed Uncertainty Quantification for Neural PDE Solvers

Abstract:

Neural PDEs offer efficient alternatives to computationally expensive numerical PDE solvers for simulating complex physical systems. However, their lack of robust uncertainty quantification (UQ) limits deployment in critical applications. We introduce a model-agnostic, physics-informed conformal prediction (CP) framework that provides guaranteed uncertainty estimates without requiring labelled data. By utilising a physics-based approach, we are able to quantify and calibrate the model's inconsistencies with the PDE rather than the uncertainty arising from the data. Our approach uses convolutional layers as finite-difference stencils and leverages physics residual errors as nonconformity scores, enabling data-free UQ with marginal and joint coverage guarantees across prediction domains for a range of complex PDEs. We further validate the efficacy of our method on neural PDE models for plasma modelling and shot design in fusion reactors.

Speaker: Vignesh Gopakumar (UKAEA)

11:20 → 11:30 Q&A

Speaker: Vignesh Gopakumar (UKAEA)