Source-linked AI summary
Probabilistic and Geometry Aware Neural Surrogate of Scrape Off Layer Plasma Simulations
Gabriele Gianuzzo, Stefan Dasbach, Fleur Hendriks, Sven Wiesen, Vlado Menkovski
TL;DR
Deterministic surrogates can miss distinct steady states near detachment, while flattened mesh vectors discard geometry. The paper combines an exact, adjacency-preserving three-view encoding with conditional flow matching; predictions split into hot and cold modes across the transition, and a synthetic check recovers both branches.
Problem
Near detachment, similar inputs can yield qualitatively different steady states, while flattened-vector surrogates discard mesh geometry and deterministic predictions can average distinct outcomes.
Source: §1 · pp. 1–2Method
The paper unrolls the curvilinear mesh into three image views with an exact inverse and trains a conditional flow-matching model to sample field outcomes.
Source: §2 · pp. 2–3, §1 · p. 2Results
Across the transition window, predictive relative spread reaches of order 200% and draws split into hot and cold populations; on synthetic data, both branches are recovered.
Source: §3 · p. 4Takeaways & Limitations
The fixed-topology representation preserves mesh adjacency and information, while a multimodal predictive distribution represents ambiguous steady states more faithfully than their mean.
Source: §4 · p. 4The representation is tied to one fixed-topology mesh, so extending it to other divertor geometries requires re-deriving strip boundaries and adjacencies by hand.
Source: §4 · p. 5
Abstract
from arXiv · showhide
Fast surrogates for tokamak boundary-plasma simulation are typically deterministic regressors mapping a global operating point to a flattened vector of cell values. Near the divertor detachment transition the steady state is not reliably single-valued. A point estimate must average over qualitatively different plasma states, and it arrives with no statement of confidence. Moreover, the flattened vector representation discards the geometric structure of the SOLPS-ITER mesh. This work addresses both problems. We unroll the curvilinear mesh into three fixed-size image tensors whose layout preserves cell adjacency and inverts exactly, letting a convolutional network act on the geometry without loss of information. A conditional flow matching model, well suited to highly sensitive systems, is then trained on this representation. The result is an efficient, scalable surrogate that captures multiple plausible outcomes even at sensitive operating points. Along a gas-puff scan, the predictive distribution splits into a hot and a cold mode across an early regime transition. A further check on synthetic data with an injected bifurcation of known size confirms the model recovers both branches rather than their average.
1 Introduction
Tokamak boundary-plasma simulations are costly, motivating fast surrogates, but near detachment similar inputs can produce qualitatively different steady states. The paper combines a probabilistic surrogate with an adjacency-preserving mesh representation to sample plausible outcomes without averaging distinct states.
Motivation
- Converged SOLPS-ITER steady states take hours to weeks, making workloads requiring hundreds of solves, such as uncertainty propagation or design search, tedious.Detachment depends on coupled plasma and neutral transport.Source: §1 · p. 1
Limitations of existing surrogates
- Existing surrogates map eight global scalars to flattened cell vectors, which discard spatial layout and struggle to capture bifurcating states near detachment.Deterministic models can average qualitatively different outcomes and require high precision and substantial training data near the critical point.Source: §1 · pp. 1–2
Contributions
- The paper introduces conditional flow matching on a lossless, adjacency-preserving encoding to sample possible steady states for a given input.The encoding represents the edge mesh as three CNN-ready tensors.Source: §1 · p. 2
2 Method
The method encodes a fixed-topology SOLPS-ITER mesh as three image views that preserve and exactly restore cell adjacency. A shared conditional flow-matching network models fields probabilistically, with sampling used to produce both predictions and per-pixel uncertainty.
Data
- The database contains 7221 converged states on one fixed-topology 104 × 50 mesh, with eight scalar inputs and 22 field channels.Channels span over ten orders of magnitude and are transformed and standardized per channel.Source: §2 · p. 2
Geometry-aware representation
- The representation partitions the mesh into six connected strips and unrolls them in adjacency-preserving order, retaining seven inter-strip adjacencies.An explicit index map stores cell-to-pixel assignments; Figure 1 marks the seven adjacencies with matching border colours.Source: §2 · p. 2, Figure 1 · p. 2
- Three canvases—top (A, B), mid (C, D, E), and bottom (F)—restore each adjacency through boundary padding and repeated feature-map reinjection.Reinjection couples the views in one forward pass and adds no parameters.Source: §2 · p. 2
- The three views share a 55 × 134 canvas, and the cell-to-pixel mapping covers all 5200 cells exactly once with an exact round trip.Evaluation uses the reconstructed mesh, so representation introduces no error.Source: §2 · p. 2
Conditional flow matching
- A shared network models each view’s field distribution with conditional flow matching, transporting a simple prior toward data along a prescribed path.The model uses data prediction with self-conditioning and recovers velocity in closed form.Source: §2 · p. 3
- At inference, lockstep views integrate the probability-flow ODE with K Euler steps; across S prior draws, per-pixel sample means on physical fields give point predictions and standard deviations quantify uncertainty.Source: §2 · p. 3
Network and training
- A two-encoder U-Net with a dedicated conditioning route and time-dependent FiLM modulation has 9.4M parameters and is trained for 200 epochs.The best-validation checkpoint is selected from a 90/10 split; Table 1 reports electron-temperature log MAE and log bias by regime.Source: §2 · p. 3, Table 1 · p. 3
3 Results
Along a gas-puff scan, the model reproduces the target-temperature response and predicts distinct hot and cold populations across the transition. Its samples track both branches in a synthetic bifurcation, while reconstruction fidelity is strongest across the three divertor regimes carrying the physics.
Evaluation
- A fifty-sample estimate for one operating point costs 79.4 s, compared with hours to weeks for one SOLPS-ITER solve.The gas-puff scan and synthetic verification use 500 prior draws for finer distributional resolution.Source: §3 · p. 3
Reconstruction fidelity
- Across the three regimes carrying divertor physics, the 2D field is reconstructed with negligible bias to a multiplicative error of factor 1.2–1.6.In cold-core, the model over-predicts the near-zero field by median factors of about two in two dimensions and three at the target, though the absolute error stays below one electronvolt.Source: §3 · p. 3
- The twenty worst of 1,465 test cases account for over half the global mean absolute error.The reported errors are set by a heavy tail.Source: §3 · p. 4
Uncertainty across a gas-puff scan
- T_OT_e falls by more than two orders of magnitude while T_OMP_e falls by about one order and plateaus near 200 eV along the gas-puff slice.The target cools far faster than upstream, the reported signature of scrape-off layer power dissipation.Source: §3 · p. 4
- Across the transition, predicted draws split into hot and cold populations with few between, while draws cluster tightly on either branch.The closest input pair among 7,221 raw simulations had puff rates differing by under 0.2% but T_OT_e values of 300 eV and 9.9 eV.Source: Figure 2 · p. 4, §3 · p. 4
Verification against a bifurcation of known size
- The synthetic bifurcation’s sampled T_OT_e recovers both exact branches across the branch-coexistence region.The synthetic check uses a bifurcation with known uncertainty, unlike the real data where the true conditional distribution is not exposed.Source: §3 · p. 4
4 Discussion and limitations
The authors argue that three image tensors can preserve mesh adjacency and information for convolutional processing, while a multimodal distribution better represents ambiguous steady states than a mean. The contribution is bounded by the lack of a deterministic-baseline accuracy comparison, the cold-core validity range, and reliance on one fixed-topology mesh.
4 Discussion and limitations
- The surrogate’s contribution concerns representing ambiguous states and mesh geometry, not demonstrating better point accuracy than the deterministic baseline.The authors report no head-to-head accuracy comparison.Source: §4 · pp. 4–5
- Cold-core target and error results are indicative only because those solutions lie outside the fluid-neutral model’s validity range.Source: §4 · p. 5
- The representation is tied to one fixed-topology mesh; extending it to other divertor geometries requires manually re-deriving strip boundaries and adjacencies.Source: §4 · p. 5