Scrape-off Layer Surrogate Training, Inference & Coupling Ecosystem (SOLSTICE)

SOLSTICE is a unified framework for deploying neural-networks surrogates of SOLPS-ITER simulations of DIII-D tokamak edge plasmas and by extension generalizing to other machines (https://github.com/iterorganization/SOLPS-ITER/). The models predict the background fluid plasma state (Te, Ti, ne, ua, heatflux, radiated power) from control parameters and the kinetic neutral source terms (Sp, Sne, Qe, Qi, Sm) from the local plasma conditions. Neural-network weights are released as self-describing bundles ready for application requiring no training code and no training data. SOLSTICE is constructed from an unstructured-first mesh design principle so that canonical cells/faces/vertices match the GOAT wide-grid SOLPS-ITER discretization, with structured grids being a special case (https://github.com/TME-TFE/goat).

Data and Resources

Field Value
accessRights Public
bgInfo https://arxiv.org/abs/2607.21407
catalogTarget ckan
citation https://doi.org/10.48550/arXiv.2607.21407
cndp FUSION
conformsTo IMAS
creationMethod Ensemble simulations were carried out by parameter scans of the SOLPS-ITER code for a baseline equilibria from DIII-D L-mode reference discharge #174310 over control conditions of input power, core density, main ion gas fueling, and radial diffusivity transport coefficients. Candidate runs for training neural-network surrogates were first verified to meet steady-state convergence criteria. Private utilities to extract and post-process simulation data were used to prepare datasets for machine learning tasks (https://github.com/ORNL-Fusion/SOLPS-routines-private). The trained surrogate model is released for public under code license Apache-2.0 (https://github.com/ORNL-Fusion/solstice). A manuscript of the scientific workflow and analysis has been prepared and submitted for publication (https://doi.org/10.48550/arXiv.2607.21407). Dataset can be converted from NetCDF to IMAS format through a Julia package (https://github.com/ProjectTorreyPines/SOLPS2imas.jl).
creatorEmail diawa@ornl.gov
creatorName Abdou Diaw
creatorWebsite https://www.ornl.gov/staff-profile/abdourahmane-diaw
dataAuthType public
doi https://doi.org/10.6084/m9.figshare.32048490
funder This research was supported in part by the U.S. Department of Energy Fusion Energy Innovation Research Engine Collaborative "Advanced Profile Prediction for Fusion Pilot Plant Design (APP-FPP)" under Award No. DE-SC0025853 and by Contract No. DE-AC05-00OR22725 at Oak Ridge National Laboratory (ORNL), managed by UT-Battelle, LLC for the Office of Science of the U.S. Department of Energy.
intendedUse https://github.com/ORNL-Fusion/solstice
lang en
lastUpdateDate 2026-08-27
libsPkgsUsed PyTorch, libEnsemble
license Apache License 2.0
modelAccessMethod https://github.com/ORNL-Fusion/solstice
ndp_creator_md5 5fab30ad0f7adaaa4470eb1d13b55190
pipelineTag simulation-prediction
pocEmail diawa@ornl.gov
pocName Abdou Diaw
pocWebsite https://www.ornl.gov/staff-profile/abdourahmane-diaw
publisherType org
purpose The SOLSTICE AI model is intended for deploying neural-network surrogates of SOLPS-ITER simulations to accelerate code coupling workflows and plasma boundary predictions for integrated modeling and control-oriented tasks. The framework will be expanded to include multiple equilibria and different experimental configurations in order to generalize the application across multiple devices for real-time, design, and operational needs.
status submitted
trainingData https://doi.org/10.6084/m9.figshare.32048490
updateFreq Annually
uploadType model
usageInfo https://github.com/ORNL-Fusion/solstice
Tags
  • SOLPS-ITER
  • edge plasma
  • machine learning
  • plasma boundary
  • scrape-off layer
  • surrogate modeling
isopen False
metadata_created 2026-10-30T16:54:21.937701
metadata_modified 2026-10-30T16:54:21.937707
notes SOLSTICE is a unified framework for deploying neural-networks surrogates of SOLPS-ITER simulations of DIII-D tokamak edge plasmas and by extension generalizing to other machines (https://github.com/iterorganization/SOLPS-ITER/). The models predict the background fluid plasma state (Te, Ti, ne, ua, heatflux, radiated power) from control parameters and the kinetic neutral source terms (Sp, Sne, Qe, Qi, Sm) from the local plasma conditions. Neural-network weights are released as self-describing bundles ready for application requiring no training code and no training data. SOLSTICE is constructed from an unstructured-first mesh design principle so that canonical cells/faces/vertices match the GOAT wide-grid SOLPS-ITER discretization, with structured grids being a special case (https://github.com/TME-TFE/goat).
num_resources 3
num_tags 6
title Scrape-off Layer Surrogate Training, Inference & Coupling Ecosystem (SOLSTICE)