Machine learning surrogates for multi-geometry EDGE2D-EIRENE studies of MAST Upgrade
Abstract
Two-dimensional edge plasma transport simulations provide a detailed description of divertor plasma behaviour, but their computational cost limits large parameter scans and iterative modelling. We present a database of 23,379 converged EDGE2D-EIRENE simulations for MAST Upgrade (MAST-U), spanning three divertor magnetic configurations, variations in separatrix density and input power, deuterium fuelling locations, and nitrogen seeding rates and locations. Independent multilayer perceptron surrogates are trained for twelve scalar outputs, including divertor front positions, target electron temperature and core-edge boundary quantities. On a held-out test set, the surrogates achieve 0.9462 ≤ R2 ≤ 0.9994, substantially outperforming linear regression baselines. Dense surrogate scans quantify the dependence of carbon radiation front motion on separatrix density, divertor geometry and fuelling location. Nitrogen seeding scans reveal a strong dependence on seeding location, with high-field-side seeding rapidly reducing core-boundary electron temperature, while divertor seeding produces little response. Batch inference for all twelve outputs at 786,432 operating points takes 1.76 s on GPU, compared with a median wall time of 2.61 hours using four CPU cores for separate EDGE2D-EIRENE simulations. These models enable rapid comparison of geometry and source-location effects on divertor and core-boundary conditions within the sampled simulation domain.
Related
- The coefficient map, not the basis, limits how machine-learning surrogates of metal–organic framework water-adsorption beds transfer to unseen materials
- Fracture-Bench: A Controlled Benchmark Suite for Machine Learning Surrogates in Phase-Field Fracture Mechanics
- Can Universal Machine-Learning Interatomic Potentials Capture Small Polarons? A Geometry-Based Benchmark in Transition-Metal Oxides
- Geometry-native machine learning reconstruction of DSMC moment fields with support monitoring
- Learning on-the-fly quasi-stationarity diagnostics from low-dimensional surrogates
BibTeX
@misc{holt2026machine,
title = {{Machine learning surrogates for multi-geometry EDGE2D-EIRENE studies of MAST Upgrade}},
author = {George K. Holt and Lorenzo Zanisi and Romain Futtersack and James Simpson},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.41756},
eprint = {2610.41756},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.41756},
}
RIS
TY - UNPB
TI - Machine learning surrogates for multi-geometry EDGE2D-EIRENE studies of MAST Upgrade
AU - George K. Holt
AU - Lorenzo Zanisi
AU - Romain Futtersack
AU - James Simpson
PY - 2026
DA - 2026/10/30
PB - arXiv
DO - 10.48550/arxiv.2610.41756
UR - https://arxiv.org/abs/2610.41756
AB - Two-dimensional edge plasma transport simulations provide a detailed description of divertor plasma behaviour, but their computational cost limits large parameter scans and iterative modelling. We present a database of 23,379 converged EDGE2D-EIRENE simulations for MAST Upgrade (MAST-U), spanning three divertor magnetic configurations, variations in separatrix density and input power, deuterium fuelling locations, and nitrogen seeding rates and locations. Independent multilayer perceptron surrogates are trained for twelve scalar outputs, including divertor front positions, target electron temperature and core-edge boundary quantities. On a held-out test set, the surrogates achieve 0.9462 ≤ R^2 ≤ 0.9994, substantially outperforming linear regression baselines. Dense surrogate scans quantify the dependence of carbon radiation front motion on separatrix density, divertor geometry and fuelling location. Nitrogen seeding scans reveal a strong dependence on seeding location, with high-field-side seeding rapidly reducing core-boundary electron temperature, while divertor seeding produces little response. Batch inference for all twelve outputs at 786,432 operating points takes 1.76 s on GPU, compared with a median wall time of 2.61 hours using four CPU cores for separate EDGE2D-EIRENE simulations. These models enable rapid comparison of geometry and source-location effects on divertor and core-boundary conditions within the sampled simulation domain.
ER -