acceptodds
Under review as a conference paper at ICLR 2027

SCOPE: Expanding Generative Coverage for STELLARATOR Design

Abstract

Stellarator design for fusion requires multiple plasma configurations that satisfy requested properties and provide alternatives for subsequent physics and engineering development. Finding these configurations is expensive, and high-quality training designs are scarce. Generative models can accelerate candidate production, creating an opportunity to expand their repertoire by learning from new physical evaluations. This introduces a mismatch between discovery and learning: once a design enters the search archive, it ceases to be novel, while the generator may still need further training to learn its region. We introduce SCOPE, a framework that separates novelty for current exploration from the continued learning value of discoveries. SCOPE starts from a conditional flow model of full plasma boundaries trained on physically screened designs. In each round, generated candidates are checked for equilibrium, request satisfaction, and physical quality, and compared against a growing archive to assess novelty. This feedback guides policy updates toward novel, useful designs, while valid discoveries retain training priority after archival. Experiments show that SCOPE generates higher-quality candidates and broadens exploration within the target design domain. It produces 2.4 times as many distinct qualified designs beyond current baseline in ordinary generation. In online search, SCOPE is able to discover more valid and qualified designs in controlled budget.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI for Fusion, generative model, reinforcement learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_8TmROpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SCOPE, a generative optimization framework that enables generative models to continuously learn from newly discovered solutions during the search process. By incorporating these discoveries into the generative policy, SCOPE aims to progressively expand the coverage of high-quality solution regions and facilitate further exploration of promising designs.

Critical strengths
  1. The paper presents an interesting perspective on the relationship between discovery and learning in generative optimization. A newly discovered solution does not necessarily imply that the generative model has adequately learned its surrounding region. By explicitly distinguishing discovery from learning, the proposed framework provides a reasonable motivation for expanding generative coverage through continued training.

  2. The paper develops a comprehensive framework that integrates exploration and continual learning. It introduces a reward mechanism to encourage exploration of previously unexplored regions, together with replay-based learning and policy refinement to progressively expand the regions accessible to the generative model.

Critical weaknesses
  1. The overall paradigm of generating candidate solutions, evaluating newly discovered solutions, incorporating them into model parameters through training, and iteratively improving exploration has been widely explored in LLM-based iterative optimization. Although SCOPE applies this paradigm to continuous generative models, it remains unclear what fundamentally new methodological contributions distinguish it from existing iterative optimization frameworks.

  2. Although the experiments demonstrate that SCOPE can discover more qualified designs under a fixed physical evaluation budget, the framework generates a large number of candidate solutions and still requires expensive physical evaluations for the selected candidates. Moreover, the experiments primarily compare SCOPE with other generative approaches, without including conventional non-generative optimization methods, particularly quality-diversity search algorithms. Consequently, it remains unclear whether SCOPE provides a meaningful advantage in search efficiency over existing optimization methods when computational costs are taken into account.

  3. Table 3 shows that SCOPE generates more distinct qualified designs than SCOPE-Uniform, but its joint success rate decreases from 59.38% to 50.00%. This suggests a potential trade-off between generation quality and diversity. It remains unclear whether SCOPE achieves broader generative coverage at the expense of generation quality, and whether this trade-off can be better balanced. A sensitivity analysis of the replay priority weight would help clarify how the proposed mechanism affects both objectives.

Questions for authors

Please see the weaknesses above.

2 comments

Thank you for the insightful comments. Below we clarify SCOPE's methodological distinction, compare search efficiency against non-generative QD, and provide the requested replay-priority sensitivity analysis.

1.[W1] Methodological distinction from LLM-based iterative optimization and archive distillation

While LLM-based iterative frameworks (e.g., FunSearch, ELM, and recent agentic stellarator optimization in Agentic Stage-One) share the high-level loop of proposal and evaluation, stellarator boundary design poses a distinct continuous-manifold challenge. In Agentic Stage-One, an LLM schedules discrete optimization stages, Fourier-mode truncations, and multi-objective proxy loss weights on top of a numerical solver—requiring repeated inner-loop MHD equilibrium solves per trajectory.

SCOPE addresses the mismatch between discovery and learning: archival novelty changes immediately, whereas incorporating a discovery into the generator requires continued training. It therefore separates current exploration rewards from persistent acquisition-time replay priority and combines them with signed feedback updates.

To isolate the benefit of this coupled online loop from post-hoc archive fitting, we compared SCOPE against two archive-distillation baselines that fine-tune the same pretrained FM for 2,048 updates on fixed valid archives (retaining quality anchors and reference regularization). Evaluated on 8 feedback requests × 32 ordinary outputs = 256 evaluations (averaging two paired seeds):

MethodJoint qualification (%)Qualified coverage (K/NK/N)
Offline FM (submitted Table 3)58.5912 / 256
Pretrained FM + QD-elite distillation38.8722 / 256
Pretrained FM + SCOPE-feedback distillation53.1336.5 / 256
SCOPE online generator50.2048 / 256

SCOPE achieves 2.18× the qualified coverage of QD-elite distillation and 31.5% more coverage than distilling its own collected feedback archive at comparable qualification, confirming the value of coupling persistent discovery weighting with online signed policy updates.

2.[W2] Comparison with non-generative QD search and computational efficiency

In stellarator design, sampling candidate boundaries from a flow model on a GPU takes seconds, whereas numerical MHD equilibrium verification (DESC; Panici et al., 2023) dominates runtime (>98% of total wall-clock cost, including candidate generation, geometric screening, physical evaluation, and policy updates). Conventional solver-based optimization must iteratively balance MHD force-balance equations against multiple competing proxy losses (quasisymmetry, magnetic well, aspect ratio, and rotational transform), making each trajectory far costlier than a single equilibrium solve. Conversely, non-generative black-box search via direct geometric perturbation avoids inner-loop solver gradients but frequently violates MHD equilibrium or request tolerances, wasting expensive evaluations.

By learning from validated feedback to refine the proposal distribution, SCOPE makes single-point physical evaluations highly productive. We compared SCOPE against domain-adapted MAP-Elites with Iso+LineDD under the exact same physical-quality gates, geometric discovery rules, and 2,048-evaluation ceiling over eight rounds (two paired runs):

MethodSearch-candidate qualification (%)Qualified new discoveriesTotal online cost (hours)Cost per qualified discovery (s)
QD search (Iso+LineDD)8.24189.971,993
SCOPE40.4837510.81104

With only 8.4% additional total time, SCOPE delivers 4.9× higher search candidate qualification (Fraction of all evaluated candidates that qualify) and 20.8× as many qualified discoveries—reducing the computational cost per qualified discovery by 19.2× (104 s vs. 1,993 s) while simultaneously producing a fast, reusable conditional generator.

3.[W3] Quality–coverage trade-off and replay sensitivity

SCOPE separates feedback for learning from qualification for reporting. New discoveries receive replay priority once they satisfy equilibrium recovery and request conditions (V=1V=1). This retains valid intermediate discoveries as learning signals: requiring every discovery to immediately satisfy all joint-quality criteria would further sparsify exploration feedback. Quality anchors continue to support qualified generation, while persistent discovery credit keeps new regions represented in training after archive admission.

Increasing β\beta gives these discoveries more replay weight. The resulting change is visible in the number of qualified outputs (GG) and the distinct qualified novel designs they contain (KK).

Replay weightQualified outputs G (%)Qualified coverage K
0 (Uniform)147 (57.42%)40
2131 (51.17%)44
4126 (49.22%)46
16131 (51.17%)53

Uniform produces more passing outputs; discovery priority produces more new qualified alternatives. At β=16\beta=16, the generator produces 131 qualified outputs containing 53 distinct novel designs, compared with 147 qualified outputs containing 40 under Uniform. Thus, the same 256 evaluations yield 13 additional distinct qualified designs (+32.5%).

This is the purpose of persistent discovery credit: to turn exploration feedback into a broader reusable repertoire. The replay weight controls how strongly learning emphasizes these discoveries, while every design counted in K satisfies the same full qualification criteria.

Reviewer_KmJsOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SCOPE, an AI system for designing magnetic devices intended to confine extremely hot plasma for fusion energy (stellarator). The authors argue that generative methods tend to reproduce already existing designing thus not discovering designs that are valid but also novel (not seen earlier). Thus, the paper makes the generative model learn iteratively from feedback with a focus on discovering novel designs instead of repeating known designs. Their experiments show that SCOPE produces more physically viable, novel designs compared to other baselines.

Critical strengths
  • The paper is well-written and studies an important problem.

  • It targets the bottleneck between discovery and generation in this particular setting.

  • The method is seen to have substantial gains over the given baselines and experiments seem to be well-thought and thorough.

  • The system combines different evaluators and metrics to build an agent-like system for solving the task at hand.

Critical weaknesses
  • This line of work resembles conceptually with well-studied line of work on LLM-based evolutionary algorithms which use LLMs as optimizers to improve discovery by exploring the search space better. I recommend looking into these works and adding them to the related work section as the work seems to draw some inspiration from evolutionary frameworks.

  • How is the reward calculated? Does it use a pareto front or is it reduced to a scalar reward? How are the tradeoffs handled?

  • How does the system handle noisy or biased rewards?

  • How reliable are the proxies used to assign the rewards? Is the system as good as the reward models?

Questions for authors

See weaknesses.

3 comments

Thank you for your positive assessment and helpful questions. We will incorporate the suggested literature on LLM-based evolutionary search and clarify our reward formulation and physical evaluation pipeline below.

1. [W1] Relation to LLM-based evolutionary optimization

We agree that FunSearch, ELM, and recent agentic stellarator workflows (Agentic Stage-One) are closely relevant: all connect model proposals with iterative evaluation, and ELM further distills evolutionary discoveries into conditional generators. We will add a dedicated discussion of these works in revision.

  • Shared reliance on a physical verification harness: In stellarator design, proposing boundary coefficients or search heuristics does not by itself establish MHD equilibrium or target physical properties. As demonstrated in Agentic Stage-One, an LLM can effectively schedule parent selection, Fourier-mode stages, objective weights, and budgets, but it still relies on a deterministic DESC harness for equilibrium solves, physical diagnostics, and acceptance checks.
  • Complementary roles of SCOPE and LLM-guided search: While LLM agents improve search planning and budget allocation across optimization runs, SCOPE solves the complementary problem of compiling expensive physical discoveries into a reusable, request-conditioned continuous generator (modeling the 168-dimensional Fourier boundary manifold via acquisition-weighted replay w=1+βmacqw = 1 + \beta m_{\mathrm{acq}} and signed flow updates). Because they operate at different layers, they naturally complement one another: SCOPE can rapidly supply diverse, physically qualified starting boundaries, LLM-guided optimization can schedule downstream local refinement, and the resulting validated designs can be fed back into SCOPE to further expand generative coverage.

2.[W2] Scalar reward structure and multi-objective trade-offs

SCOPE uses a validity-gated scalar reward (Eq. 2 and Appendix A.4, Eq. 8):

r=m+0.15s+0.05U(V=1),r=−1(V=0).r = m + 0.15s + 0.05U \quad (V=1), \qquad r = -1 \quad (V=0).

Explicit priorities in learning. Validity (V=1V=1) requires geometric admissibility, numerical MHD equilibrium recovery, and satisfaction of the requested tolerances. Among valid designs, a distinct discovery (m∈{0,1}m\in\{0,1\}) receives a bonus of 1.0. Progress toward the geometric novelty thresholds (s∈[0,1]s\in[0,1]) and bounded physical utility (U∈[0,1]U\in[0,1]) contribute at most 0.20 combined. The physical utility equally combines a bounded QS-quality score and the fraction of radial nodes with positive magnetic well. This scalarization prioritizes discovering new valid regions while retaining graded incentives for novelty progress and physical quality.

Physical qualification of reported coverage. The learning reward and the qualification criterion serve distinct roles. Every design counted in qualified coverage (Gτ=1G_\tau=1) must satisfy V=1V=1, Q<10−3Q<10^{-3}, and nonnegative magnetic well at all 50 off-axis radial nodes. Thus, the reward allocates learning effort among valid candidates, while reported coverage measures additional distinct designs meeting the same fixed physical-quality standard.

3.[W3] Handling numerical variation and extreme feedback

To ensure stable policy updates across diverse requests and numerical solves:

  1. Within-request normalization and clipping: Rewards are standardized per request (with standard-deviation floor σmin⁡=0.1\sigma_{\min} = 0.1) and clipped to [−2,2][-2, 2] (Eq. 3), preventing outlier evaluations from dominating gradients.
  2. Trust-region policy updates: We bound per-sample matching-error changes via PPO-style clipping (κ=0.2\kappa = 0.2, Eq. 5), anchor velocity fields to the initial policy (λref=0.1\lambda_{\mathrm{ref}} = 0.1, Eq. 9), and monitor velocity drift every 16 steps with automatic rollback if drift exceeds 0.020.02 (Appendix A.4).

4.[W4] First-principles physical evaluation (no learned reward proxies)

Crucially, SCOPE does not use learned surrogate or neural reward models; every candidate is evaluated directly by the first-principles numerical MHD equilibrium solver DESC (Panici et al., JPP 2023), eliminating neural reward hacking:

  • Equilibrium certification: Every accepted equilibrium must achieve normalized force-balance RMS ≤10−4\le 10^{-4} on both working and denser validation grids with nested flux surfaces and positive Jacobian (Appendix A.3). Candidates that fail to converge within the fixed 20-iteration solver budget receive r=−1r = -1, steering the generator away from ill-conditioned boundaries.
  • Validated physical metrics: Quasisymmetry (QS) is computed from symmetry-breaking Boozer Fourier harmonics across 10 radial surfaces (Landreman & Paul, PRL 2022)—a metric experimentally validated to reduce neoclassical transport (Canik et al., PRL 2007)—while the magnetic well profile W(ρ)W(\rho) is verified across 50 off-axis radial nodes to screen for ideal-MHD interchange stability (Landreman & Jorge, JPP 2020).
Reviewer_KmJsOpenReview ↗

Follow up on Author Response

Thank you for your response. Based on the responses, I choose to retain my score.

Reviewer_CJvnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes SCOPE - a framework for generating diverse stellarator plasma configurations that satisfy desired design specifications and physical quality criteria. It starts with a conditional flow model trained on plasma boundaries, then improves the generator through repeated physical evaluation and policy updates. The main idea is to reward novel designs during exploration while continuing to prioritize previously discovered designs during training. On target-domain requests, SCOPE produces 102 distinct qualified designs compared with 42 for Diffusion for Fusion. Controlled experiments also show improvements in discovery and final generative coverage, although the coverage advantage is weaker on new requests.

Critical strengths
  1. Interesting setup for scientific discovery: I like the distinction between finding a new design and learning to generate its neighborhood. Continuing to train on discoveries after they lose their novelty is well motivated.

  2. Generated designs undergo numerical equilibrium checks and explicit physical-quality evaluation. This makes the results more meaningful than evaluating geometric diversity alone.

  3. Strong results within the target domain: SCOPE substantially improves qualified-design generation and produces approximately 2.4 times as many distinct qualified designs as DFF.

  4. The experiments include useful controls, replay ablations, a second online seed, and sensitivity analyses for novelty thresholds and counting order.

Critical weaknesses
  1. My main concern is generalization to new design requests. The strongest coverage results use a development panel that includes the requests used for online feedback. On 64 additional requests, SCOPE produces 88 distinct qualified designs, fewer than DFF’s 97 and DFF-PCA-Joint’s 109. These requests also remain within the same stellarator design setting. How does the approach perform outside the training range or with different physical settings and field-period configurations? The results support improved quality on new requests more clearly than expanded coverage.

  2. The experiments need to include comparisons with stronger baselines. SCOPE combines novelty rewards, replay, and signed flow-matching updates, but its distinction from existing quality-diversity and archive-distillation methods is unclear. Comparisons with these methods under the same physical-evaluation budget would help establish its added value.

  3. The practical impact of the additional geometric coverage needs further validation. Equilibrium recovery, quasisymmetry, and magnetic-well checks are useful early-stage criteria, but do not establish whether the additional designs offer better options for subsequent engineering. A more comprehensive analysis of downstream refinement and confinement would better demonstrate the benefit of the expanded repertoire.

Questions for authors

Please address the commnets in the weaknesses section.

2 comments

Thank you for the constructive review. We added cross-backbone request-transfer experiments, non-generative QD and archive-distillation baselines, and Stage-2 modular coil designs to address each of your points.

1. [W1]Generalization to additional requests and backbone choice

To test whether the request-transfer coverage gap in Table 2 stems from SCOPE's online learning rule or from the offline starting generator, we applied SCOPE to both our full-boundary Flow Model (FM) and the exact PCA50-DDPM backbone used by DFF (Padidar et al., 2025). Both runs use the same eight feedback requests, online schedule, physical-evaluation ceiling, and geometric novelty criteria (converting Eqs. 5–6 to noise-prediction error for DDPM). We evaluate zero-shot transfer on the 64 additional requests × 16 ordinary outputs = 1,024 evaluations (averaging two paired seeds):

GeneratorRefinementJoint qualification (%)Qualified coverage (K/NK/N)
FMNone62.8971 / 1,024
FMSCOPE61.4780.5 / 1,024 (+13.4%)
PCA50-DDPMNone17.29116 / 1,024
PCA50-DDPMSCOPE45.56 (2.6×)225 / 1,024 (+94.0%)
  • SCOPE consistently expands transfer coverage across backbones: Applying SCOPE to PCA50-DDPM nearly doubles distinct qualified designs on unseen requests (116 →\to 225) while boosting joint qualification by 2.6× (17.29% →\to 45.56%). These cross-backbone results show that SCOPE improves coverage on additional requests with both generators and can build effectively on the broader initial repertoire of PCA50-DDPM.
  • Why instantiate SCOPE with FM? Each evaluated boundary requires a costly numerical MHD equilibrium solve, so starting from FM's much higher physical qualification (62.89% vs. 17.29%) makes the online evaluation budget substantially more productive. FM also provides a unified velocity-matching objective across pretraining, replay, and signed feedback refinement. Extending beyond the two-field-period (NFP=2N_{\mathrm{FP}}=2) setting changes the underlying Fourier mode topology and naturally pairs with multi-NFPN_{\mathrm{FP}} pretraining in future work.

2. [W2]Comparison with quality-diversity (QD) search and archive distillation

We implemented domain-adapted non-generative MAP-Elites with Iso+LineDD under the exact same physical-quality gates, geometric discovery rule, and budget of 2,048 physical evaluations over eight rounds (two paired seeds):

  • Online search productivity: Direct QD search yields 18 qualified new discoveries (8.24% qualification), whereas SCOPE achieves 375 qualified discoveries (20.8× QD) at 40.48% qualification.
  • Archive distillation vs. SCOPE: We also compared SCOPE against distilling fixed valid archives (Pretrained FM + QD-elite distillation and Pretrained FM + SCOPE-feedback distillation), fine-tuning the same offline FM for 2,048 updates while retaining quality anchors and reference regularization. On 8 feedback requests × 32 ordinary outputs = 256 evaluations:
MethodJoint qualification (%)Qualified coverage (K/NK/N)
Offline FM (submitted Table 3)58.5912 / 256
Pretrained FM + QD-elite distillation38.8722 / 256
Pretrained FM + SCOPE-feedback distillation53.1336.5 / 256
SCOPE online generator50.2048 / 256

At comparable qualification (~50%), SCOPE achieves 2.18× the qualified coverage of QD-elite distillation and 31.5% more coverage than offline distillation of its own collected archive. Simply fitting accumulated discoveries post hoc does not recover the repertoire gained by coupling novelty-driven search with persistent acquisition-priority replay and signed advantage updates.

3.[W3] Additional Stage-2 coil design and equilibrium-recovery assessment

Stellarator design commonly proceeds from plasma-equilibrium optimization (Stage-1) to coil design for selected targets (Stage-2). Additional coil and engineering constraints can favor different plasma boundaries, motivating alternative targets and subsequent refinement (Jorge et al., 2023). Appendix E illustrates this role through two configurations satisfying the same request but offering complementary QS–magnetic-well trade-offs.

We fitted 16 modular filament coils to each configuration under common optimization settings. Preserving the original finite-pressure targets and accounting for plasma currents through virtual casing, the fits achieve normalized RMS normal-field mismatches of 0.308% and 0.289% on an independent validation grid.

We then performed finite-pressure free-boundary calculations, holding the fitted coils, coil currents, plasma profiles, and toroidal flux fixed while allowing the plasma boundary to respond. Case A passes the paper's MHD-recovery checks, retains a QS residual of 4.32×10−44.32\times10^{-4}, and preserves nonnegative magnetic well at both 50 and 100 radial nodes. Case B does not pass the same recovery checks under the tested settings: its force-balance RMS is 1.32×10−41.32\times10^{-4} on the working grid and 1.11×10−41.11\times10^{-4} on the denser grid, exceeding the 10−410^{-4} threshold. Separately, the final normalized normal-field/magnetic-pressure RMS residuals are 0.616%/0.612% for A and 0.449%/0.528% for B.

Case A provides a concrete downstream continuation of a SCOPE-generated configuration: with the fitted coils held fixed, the free-boundary calculation passes our MHD-recovery checks and retains the required QS and magnetic-well properties. The different outcomes of the two Appendix E configurations illustrate how the expanded repertoire supplies alternative targets for downstream coil design, selection, and refinement.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.