Speaker
Description
Edge plasma simulations with SOLPS-ITER are expensive and sensitive to initialization, especially when spanning wide parameter ranges or tailoring to specific discharges. We present a reduced-modeling workflow that combines (i) warm starts via nearest-neighbor initialization from a KD-tree [1] of converged runs and (ii) learned surrogates for rapid prediction. The KD-tree selects a converged neighbor whose terminal state seeds the new run, reducing runtime relative to cold starts. For prediction, we train two complementary families of surrogates on DIII-D datasets [2] spanning gas puff, core density, cross-field transport, and diffusivities: (1) one-dimensional profile surrogates at the outer midplane and targets, using an ensemble fully connected network that achieves high R2 with typical relative errors ≲ 20% on held-out cases; and (2) two-dimensional map surrogates for poloidal fields and sources, using a U-Net [3] for pixel-aligned regressions and a variational autoencoder [4] to learn a low-dimensional latent representation that enables fast sampling and anomaly detection. To limit costly sampling in low-value regions, we employ an adaptive sampling approach [5] that targets high-variance areas and iteratively retrain the models. Finally, we demonstrate the initial inline use of the surrogate within the SOLPS-ITER loop to reduce selected EIRENE calls under physics-based constraints. The combined approach accelerates scenario generation, reduces convergence risk, and enables real-time inference in edge-plasma studies.
[1] J. L. Bentley, Commun. ACM 18, 509–517 (1975).
[2] J. D. Lore et al., Nucl. Fusion 63, 046015 (2023).
[3] O. Ronneberger et al., Lect. Notes Comput. Sci. 9351, 234–241 (2015).
[4] R. Anirudh et al., Proc. Natl. Acad. Sci. U.S.A. 117, 9741–9746 (2020).
[5] A. Diaw et al., Nat. Mach. Intell. 6, 568–577 (2024).