Pith. sign in

REVIEW 4 major objections 4 minor 30 references

A permutation-equivariant neural operator trained on ten-satellite swarms plans collision-aware maneuvers for 1,000-satellite swarms in a single forward pass and a Gauss–Newton finish.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 00:43 UTC pith:POB7MJCB

load-bearing objection Valuable operator-learning result, but the architecture can't see other agents, so the 'swarm coordination' claims should be trimmed. the 4 major comments →

arxiv 2608.00320 v1 pith:POB7MJCB submitted 2026-07-31 cs.LG cs.MAcs.SYeess.SY

Neural operator learning for collision-aware trajectory planning of spacecraft swarms

Sidhdharth D. Sikka , Suyi Gao , Zehui Lu , Rongjie Lai , Shaoshuai Mou This is my paper
classification cs.LG cs.MAcs.SYeess.SY MSC 68T0770M20
keywords Neural operatorsOperator learningDistributional controlMulti-agent systemsCollision-aware planningSpacecraft swarmsOrbital autonomy
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper claims that collision-aware trajectory planning for an entire spacecraft swarm can be collapsed into a single forward pass of a permutation-equivariant neural operator, with a batched Gauss–Newton step closing each trajectory onto exact two-body dynamics. Trained on swarms of up to ten spacecraft using only self-supervised physics losses and adversarial debris placed on its own rollouts—no optimal-trajectory labels—the operator transfers zero-shot to swarms of 1,000 amid a catalog of more than 11,000 objects, matching a per-agent optimal-control solver's terminal accuracy at comparable fuel cost while avoiding close approaches the debris-blind baseline cannot. If true, this would change the scaling of collision avoidance in congested low Earth orbit from per-pair optimization to a single amortized inference, with run time set by mission duration rather than swarm size.

Core claim

The central claim is that a single permutation-equivariant neural operator, trained on small swarms, learns a sample-invariant map from initial spacecraft, target, and debris distributions to collision-aware trajectories for the whole swarm, and that this map remains valid at swarm sizes two orders of magnitude beyond training. The paper establishes this by training the operator with physics-grounded objectives—terminal accuracy, a fuel surrogate derived from Gauss variational equations, and closest-point-of-approach penalties—plus adversarial debris that intersect the model's own nominal rollouts. At deployment, a batched Gauss–Newton finish, formulated as a 6×6 system via the Woodbury iden

What carries the argument

The load-bearing object is a permutation-equivariant, time-conditioned neural operator built from two parallel multi-head cross-attention streams—one aligning agents with target orbits, one aligning them with debris—so each agent's trajectory prediction is conditioned on both goal and obstacle distributions. A physics-informed baseline (linear or wrapped interpolation for five orbital elements plus a Keplerian true-anomaly rate) carries the dynamics, and the network learns only residuals. Training uses a closest-point-of-approach penalty with analytic within-interval separation, a fuel surrogate from the pseudoinverse of the Gauss variational equations, and adversarial debris generated as cr

Load-bearing premise

The load-bearing assumption is that a permutation-equivariant attention mechanism trained on swarms of up to ten spacecraft learns the true pairwise interaction rule, so the same rule remains valid for swarms of 1,000; the paper states this extrapolation is observed empirically and not theoretically characterized.

What would settle it

Run the trained operator on intermediate swarm sizes (N=20, 50, 200) with adversarial debris generated against the operator's own warm-started trajectories. If per-spacecraft proximity or terminal error degrades abruptly as N crosses 10—rather than the gradual, bounded degradation reported at N=1000—the size-invariant interaction rule would be refuted; alternatively, a single N=1000 trial where a newly drawn debris object intersects the finished trajectory at less than 100 m with no agent–agent cause would show the safety claim does not hold under distribution shift.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Zero-shot scaling: trained on N≤10, the finished planner maintains bounded terminal error and low proximity rates at N=100 and N=1000 over the full 11,000-object catalog.
  • Adversarial robustness: one debris object placed on each method's own nominal path strikes the debris-blind baseline on 99.3–99.8% of maneuvers, while the operator-warm finish clears it on essentially every maneuver (≤0.21% at N=1000).
  • Complexity shift: both inference and the Gauss–Newton finish have run times set by trajectory length, not agent or debris count, so the full swarm replans in under a minute at the 6-hour horizon and neither stage scales combinatorially.
  • Label-free training: the operator is trained without optimal-trajectory labels, using self-supervised physics losses and adversarial scenario generation, so the approach does not require a precomputed solution library.
  • Dynamic feasibility with preserved avoidance: the Gauss–Newton finish closes each rollout onto exact two-body dynamics and tracks the operator's raw collision-avoidance geometry within a few tenths of a percent, so accuracy and safety survive the dynamics-closing step.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same division of labor—learned collision geometry plus a certified numerical finish—should transfer to other multi-agent planning problems where pairwise constraints multiply, such as airspace deconfliction or underwater vehicle swarms; the paper notes the recipe is not specific to astrodynamics.
  • Because the interaction penalties are soft and the finish carries no collision term, the method offers no worst-case guarantee; a natural next step is a verification layer or hard-constraint projection that preserves the operator's speed.
  • The reported duration generalization is limited to the trained 1–12 hour horizon, and the paper explicitly frames receding-horizon execution as untested; closing the loop with periodic replanning and measuring convergence and safety over multiple cycles is a concrete deployment test.
  • Residual agent–agent proximity at N≥100 suggests that the training's clustered-start construction, not the architecture alone, determines internal deconfliction; varying conflict geometry in training could test whether the operator learns a general deconfliction rule or one tailored to the manufactured encounters.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes a two-stage planner for spacecraft-swarm trajectory planning: a self-supervised, permutation-equivariant transformer operator that maps initial spacecraft, target, and debris distributions to orbital-element trajectories, followed by a batched Gauss–Newton finish that closes each trajectory onto exact two-body dynamics. Training uses physics-informed losses (fuel surrogate, terminal/initial matching, and a closest-point-of-approach collision penalty) with adversarially generated debris, and no optimal-trajectory labels. The authors report that a model trained on N≤10 spacecraft generalizes zero-shot to N=1000 amid a catalog of more than 11,000 debris objects, matches a per-agent IPOPT solver at N=1, substantially reduces debris close approaches, and reduces agent-agent proximity several-fold relative to a cold-started baseline.

Significance. If the claims held in full, the paper would make a useful contribution: amortized, scalable collision-aware planning with no optimal-trajectory labels is an appealing direction, and the two-stage learning-plus-numerics division of labor is well motivated. Strengths include a self-supervised training objective that does not require precomputed optimal solutions, a CPA-based scoring method that is shown to be grid-independent, a batched Gauss–Newton finish with a conserved-vector residual that avoids element-angle singularities, evaluation on real TLE-derived initial conditions, and an unusually candid list of limitations. However, the central 'swarm collision avoidance' claim is undermined by the architecture, which contains no agent-agent attention, and the adversarial evaluation is partly self-referential because the threats are generated against the model's own rollouts and the evaluation metric is the same CPA quantity minimized in training. The significance of the paper therefore depends on how much of the claimed multi-agent coordination is real versus an emergent artifact of per-agent debris avoidance.

major comments (4)
  1. [Methods, 'Dual Cross-Attention Stack', Eqs. (24)–(26)] The architecture has no agent–agent interaction. In Eqs. (24)–(26) the spacecraft queries X_q are updated only by cross-attention to the target set H1 and the debris set Hd; there is no self-attention or pairwise term among the N spacecraft. Consequently, the trajectory of agent i is conditionally independent of the states of all other agents given its own initial state, the target set, and the debris set. The paper nevertheless claims that the operator 'applies the same permutation-equivariant interaction rule across arbitrary numbers of sample points' (Discussion) and that it 'reduc[es] proximity within the swarm several-fold' (Abstract; Results, 'Collision avoidance'). These claims are unsupported: the reported agent–agent proximity reduction (0.07–0.22% vs 0.51–0.67% at N=1000 in Table 1) cannot be attributed to learned coordination, since the model cannot even represent pairwise int
  2. [Methods, 'Adversarial debris generation'; Eq. (7)–(8); Results, Table 1] The evaluation of 'worst-case threat' evasion is partly self-referential. Adversarial debris objects are generated against the model's own debris-unaware rollout (Methods, 'Adversarial debris generation'), and the headline proximity metric is the same CPA quantity (Eq. (7)) that is penalized during training (Eq. (8)). Thus the result that the operator 'evades worst-case threats that a debris-blind baseline cannot' is partly a check that the model optimized its training objective, not a test against an independent worst-case adversary. To support the robustness claim, the authors should evaluate against threats not generated by the model itself—for example, objects placed on the nominal trajectory of a different baseline, or real close-approach events from the TLE catalog—and report a metric not used in training, such as the actual minimum separation from a high-fidelity propagation.
  3. [Table 1; 'Runtime and scaling', Fig. 5] Statistical and runtime reporting is incomplete. Table 1 reports only medians over 500 Monte Carlo trials; no error bars, confidence intervals, or interquartile ranges are given for terminal error, fuel cost, or proximity rates. Given the stochasticity of both training and scenario sampling, the claim that GNw 'degrades gradually rather than abruptly' to N=1000 is not supported by point estimates alone. In addition, Fig. 5 measures runtime at M=1000 debris, whereas the headline N=1000 results in Table 1 are reported 'amid the full >11,000-object catalog.' The runtime of the full pipeline at the actually evaluated scale is therefore not reported. The authors should provide variance measures for Table 1 and runtime measurements at the catalog scale used in the main results.
  4. [Results, 'Adversarial and debris-field trajectory planning'; Table 1] The comparison to IPOPT is limited to N=1 in the adversarial scenarios. The Abstract states that the operator matches 'a per-agent optimal-control solver's accuracy', but the only IPOPT results are single-agent cases with a single debris object. This is a narrow basis for the claim. Within the training distribution (N≤10, M small) IPOPT should be tractable and would provide a much more meaningful accuracy and fuel-cost comparison. The authors should report IPOPT comparisons for at least N=10 in both adversarial and (small-M) debris scenarios, or explicitly restrict the optimal-control-matching claim to N=1.
minor comments (4)
  1. [Data and code availability] The availability statement says model weights, processed ephemeris, and code 'will be deposited' and are 'available to editors and reviewers on request'. For a reproducibility-conscious journal, a permanent repository link or DOI should be provided in the manuscript, or at least a clear statement of the intended repository name and access date.
  2. [Discussion, limitations paragraph] The Discussion honestly notes that extrapolation is 'observed empirically but not theoretically characterized', that penalties are soft and there is no worst-case guarantee, that dynamics are deterministic two-body without J2/drag/uncertainty, and that receding-horizon execution is not evaluated. These are appropriate scope statements, but they should be reflected more prominently in the Abstract, which currently presents zero-shot generalization and collision awareness without these qualifications.
  3. [Results, 'Grid-independent proximity scoring'; Eq. (7)] The CPA refinement is a nice contribution. However, the training loss uses the same CPA as the evaluation metric; the grid-independence experiment in Fig. 4 would be stronger if it also showed that the learned policy trained with the CPA surrogate transfers to a metric based on full high-fidelity propagation (e.g., with J2), since the authors acknowledge that operational catalogs have uncertainty comparable to the 100 m threshold.
  4. [Methods, 'Scenario Sampling'] The scenario sampling is clear, but the claim that targets 'form a cluster mirroring the start cluster' is confusing when the same deviation vector is applied to every agent's own P0; please clarify explicitly whether the targets are spatially converging or simply shifted copies of the initial cluster.

Circularity Check

2 steps flagged

Partial circularity: proximity and adversarial-evasion results are measured on the same CPA quantity and threat construction used in training; zero-shot swarm-size and dynamics claims remain independently supported.

specific steps
  1. fitted input called prediction [Methods 'Interaction penalties', Eq. (8); Methods 'Loss Weights', Table 5; Results 'Collision avoidance']
    "The spacecraft–spacecraft safety radius r_s = 100 m applies only for t≥0.2T ... The spacecraft–spacecraft radius matches the 100 m threshold used in evaluation. ... Proximity is reported as a per-spacecraft rate, the percentage of planned maneuvers that pass within 100 m of another agent or debris object."

    The model is trained by minimizing J_CPA with exactly the 100 m spacecraft-spacecraft radius that defines the evaluation metric. The headline 'reducing proximity within the swarm several-fold' is therefore a report of the fitted objective value measured with the same CPA function, not an independent prediction. The N=1000 extrapolation is not forced because the model was not trained at that size, but the proximity comparison is self-referential by construction.

  2. self definitional [Abstract; Methods 'Adversarial debris generation'; Results 'Adversarial and debris-field trajectory planning']
    "combining self-supervised physics objectives with adversarial threats generated against its own rollouts. ... We construct this worst case by seeding one such object on each method’s own debris-unaware path. ... A planner blind to it is struck almost every time (GNc, 99.3–99.8% of maneuvers at all sizes)."

    The 'worst-case threat' is defined as an object placed on the planner's own debris-unaware predicted path. A debris-blind baseline follows that path by construction, so GNc's 99.3–99.8% failure is a definitional consequence of the threat construction rather than an empirical discovery. The operator's success still requires genuine debris conditioning, but the 'threats that a debris-blind baseline cannot evade' comparison is partly self-referential, and the evaluation uses the same crossing-orbit threat distribution used in training.

full rationale

The central zero-shot claim—training on N≤10 and transferring to N=1000 amid the full TLE catalog—is not circular: it is an empirical extrapolation test on real ephemeris data, and the operator does not see N=1000 trajectories during training. The Gauss–Newton finish is independently derived and batched, and IPOPT provides an external fuel/accuracy benchmark at N=1. The Huang et al. [27] framework is a self-citation (Lai is a co-author), but it is a published architecture reference, not an imported uniqueness theorem, so it is not load-bearing circularity. The partial circularity lies in two evaluation choices: the proximity metric is the same CPA penalty minimized in training (same 100 m threshold), and the adversarial threats are generated from the model's own debris-unaware rollouts, making the debris-blind baseline's failure definitional. These issues weaken the 'evading worst-case threats' and 'reducing proximity' phrasings, but they do not reduce the core operator-learning derivation to its inputs. The paper itself honestly notes that extrapolation is 'observed empirically but not theoretically characterized,' which is a limitation rather than a circular step.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

The central claims rest on a simplified deterministic two-body model, a self-consistent CPA metric, an empirically assumed sample-invariant attention operator, and hand-chosen penalty weights, radii, and adversarial constructions. No new physical entities are introduced.

free parameters (8)
  • Collision penalty weight λ_I = 10^3
    Hand-chosen weight in Eq. (12); dominates training and directly shapes avoidance behavior. No sensitivity study is reported.
  • CPA hinge scale κ = 10^4
    Scales the quadratic hinge in Eq. (8); hand-chosen, affects gradient magnitude near close approaches.
  • Safety radii (rd, rs) = 1 km debris, 100 m agent-agent
    Hand-set thresholds. Evaluation uses the same rs=100m as training, so reported proximity rates depend on this choice.
  • Adversarial offset and rotation ranges = 0.35–0.65 rd; 20–75 deg
    Defines the constructed 'worst-case' threat geometry; hand-chosen to yield well-conditioned avoidance gradients.
  • Loss weights λ_f, λ_T, λ_S = 1e-2, 1e2, 1e2
    Hand-chosen balance in Eq. (12); no sensitivity analysis.
  • Integration timestep Δt = 120 s
    Used for rollout, CPA intervals, and runtime; affects discretization, collision-scoring fidelity, and training cost.
  • Maneuver magnitudes = 1% and 10% per element
    Defines minor/major scenario classes; makes fuel costs directly comparable across trials.
  • GN fuel regularizer and trust-region schedule = λ_f small default; μ down 0.5, up 4; ≤25 iterations
    Hand-chosen finish parameters that trade terminal accuracy against control energy.
axioms (6)
  • domain assumption Unperturbed two-body Keplerian dynamics (GVE) with unpowered debris
    All training and evaluation use Eq. (1). Discussion acknowledges no J2, drag, solar radiation pressure, or third-body perturbations.
  • domain assumption Deterministic dynamics with no state-estimation uncertainty or thrust execution error
    Acknowledged in Discussion; catalog uncertainty can be comparable to the 100–500 m thresholds considered.
  • ad hoc to paper CPA linear-relative-motion approximation is a faithful collision metric
    Eq. (7) is used both in the training loss (8) and as the evaluation metric; the grid-sweep validation is internal to the same model, not against high-fidelity propagation.
  • ad hoc to paper Permutation-equivariant attention learned on N≤10 extrapolates to N=1000
    Central to the zero-shot claim; the paper states the extrapolation is observed empirically but not theoretically characterized.
  • domain assumption GVE pseudoinverse control inference gives a meaningful fuel surrogate
    Fuel loss (13) uses g†; assumes the control-affine GVE inversion is well-posed on predicted rollouts.
  • domain assumption TLE/SGP4 states are ground-truth initial conditions and debris fields
    All scenarios draw from the Space-Track TLE snapshot; SGP4 propagation is assumed accurate for initial conditions.

pith-pipeline@v1.3.0-alltime-deepseek · 16530 in / 18804 out tokens · 171877 ms · 2026-08-04T00:43:41.375368+00:00 · methodology

0 comments
read the original abstract

Autonomous spacecraft swarms must plan fuel-efficient, collision-free maneuvers in increasingly congested orbits, yet classical trajectory optimization scales poorly as pairwise safety constraints multiply with swarm size, and learning-based planners rarely transfer across swarm sizes or debris densities. Here we introduce a permutation-equivariant neural operator that maps distributions of spacecraft, targets and debris to collision-aware trajectories for an entire swarm in a single forward pass, paired with a batched Gauss-Newton finish that enforces exact orbital dynamics. The operator is trained without optimal-trajectory labels, combining self-supervised physics objectives with adversarial threats generated against its own rollouts. Trained on ten spacecraft, it generalizes zero-shot to swarms of 1,000 amid more than 11,000 catalogued objects, matching a per-agent optimal-control solver's accuracy, evading worst-case threats that a debris-blind baseline cannot, and reducing proximity within the swarm several-fold. Physics-grounded operator learning thus offers a fast, scalable alternative to optimal control for crowded orbits.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 1 linked inside Pith

  1. [1]

    Technical report, ESA/ESOC, Darmstadt, Germany (2026)

    European Space Agency Space Debris Office: ESA’s Annual Space Environment Report. Technical report, ESA/ESOC, Darmstadt, Germany (2026)

    2026

  2. [2]

    Space.com

    Space.com: Every SpaceX Starlink satellite has to dodge a collision almost weekly, and experts fear the worst. Space.com. Reporting SpaceX’s semi-annual orbital- safety filing to the FCC, covering December 2025–May 2026 (2026)

    2025

  3. [3]

    Journal of Spacecraft and Rockets60(6), 1828–1845 (2023)

    Lee, H.W., Ho, K.: Regional constellation reconfiguration problem: Integer linear programming formulation and Lagrangian heuristic method. Journal of Spacecraft and Rockets60(6), 1828–1845 (2023)

    2023

  4. [4]

    Journal of Guidance, Control, and Dynamics46(8), 1548–1563 (2023)

    Basu, H., Pedari, Y., Almassalkhi, M., Ossareh, H.R.: Computationally efficient collision-free trajectory planning of satellite swarms under unmodeled orbital perturbations. Journal of Guidance, Control, and Dynamics46(8), 1548–1563 (2023)

    2023

  5. [5]

    Journal of Guidance, Control, and Dynamics40(7), 1541–1566 (2017)

    Eren, U., Prach, A., Ko¸ cer, B.B., Rakovi´ c, S.V., Kayacan, E., A¸ cıkme¸ se, B.: Model predictive control in aerospace systems: Current state and opportunities. Journal of Guidance, Control, and Dynamics40(7), 1541–1566 (2017)

    2017

  6. [6]

    Aerospace Science and Technology 151, 109273 (2024)

    Chen, R., Dong, M., Bai, Y., Zhao, Y., Chen, X.: Trajectory planning and control 25 of spacecraft avoiding dynamic debris swarm. Aerospace Science and Technology 151, 109273 (2024)

    2024

  7. [7]

    In: Robotics Research

    van den Berg, J., Guy, S.J., Lin, M., Manocha, D.: Reciprocal n-body collision avoidance. In: Robotics Research. Springer Tracts in Advanced Robotics, vol. 70, pp. 3–19. Springer, Berlin, Heidelberg (2011)

    2011

  8. [8]

    IF AC-PapersOnLine56(2), 547–552 (2023)

    Pedari, Y., Basu, H., Ossareh, H.R.: A novel framework for trajectory plan- ning and safe navigation of satellite swarms. IF AC-PapersOnLine56(2), 547–552 (2023)

    2023

  9. [9]

    International Journal of Aeronautical and Space Sciences26(1), 314–326 (2025)

    Jung, I., Chung, D.: Genetic algorithm-based approach for improving temporal resolution in constellation operation of national satellites. International Journal of Aeronautical and Space Sciences26(1), 314–326 (2025)

    2025

  10. [10]

    Acta Astronautica 224, 325–337 (2024)

    Xu, L., Zhang, G., Qiu, S., Cao, X.: Reinforcement learning-based multi-impulse rendezvous approach for satellite constellation reconfiguration. Acta Astronautica 224, 325–337 (2024)

    2024

  11. [11]

    In: 2025 IEEE Aerospace Conference, pp

    Kuhl, W., Wang, J., Eddy, D., Kochenderfer, M.J.: Markov decision processes for satellite maneuver planning and collision avoidance. In: 2025 IEEE Aerospace Conference, pp. 1–9. IEEE, Big Sky, MT (2025)

    2025

  12. [12]

    Nature Machine Intelligence8(6), 955–968 (2026)

    An, X., Luo, S., Zhang, H., Yang, Q., Ma, Y., Wang, B., Du, J., Wang, Q.: Autonomous navigation of intelligent microrobotic swarms in unknown environments. Nature Machine Intelligence8(6), 955–968 (2026)

    2026

  13. [13]

    SpringerBriefs in Mathematics

    Bensoussan, A., Frehse, J., Yam, P.: Mean Field Games and Mean Field Type Control Theory. SpringerBriefs in Mathematics. Springer, New York (2013)

    2013

  14. [14]

    Sensors22(14), 5437 (2022)

    Wang, G., Yao, W., Zhang, X., Li, Z.: A mean-field game control for large-scale swarm formation flight in dense environments. Sensors22(14), 5437 (2022)

    2022

  15. [15]

    In: Advances in Neural Information Processing Systems, vol

    Guo, X., Hu, A., Xu, R., Zhang, J.: Learning mean-field games. In: Advances in Neural Information Processing Systems, vol. 32. Curran Associates, Inc., Red Hook, NY (2019)

    2019

  16. [16]

    Proceedings of the National Academy of Sciences117(17), 9183–9193 (2020)

    Ruthotto, L., Osher, S.J., Li, W., Nurbekyan, L., Fung, S.W.: A machine learning framework for solving high-dimensional mean field game and mean field control problems. Proceedings of the National Academy of Sciences117(17), 9183–9193 (2020)

    2020

  17. [17]

    Preprint at https://arxi v.org/abs/2205.12944 (2022)

    Lauri` ere, M., Perrin, S., P´ erolat, J., Girgin, S., Muller, P., ´Elie, R., Geist, M., Pietquin, O.: Learning in Mean Field Games: A Survey. Preprint at https://arxi v.org/abs/2205.12944 (2022)

    Pith/arXiv arXiv 2022

  18. [18]

    Journal of Computational Physics378, 26 686–707 (2019)

    Raissi, M., Perdikaris, P., Karniadakis, G.E.: Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics378, 26 686–707 (2019)

    2019

  19. [19]

    Nature Reviews Physics3(6), 422–440 (2021)

    Karniadakis, G.E., Kevrekidis, I.G., Lu, L., Perdikaris, P., Wang, S., Yang, L.: Physics-informed machine learning. Nature Reviews Physics3(6), 422–440 (2021)

    2021

  20. [20]

    Nature Machine Intelligence7(6), 954–966 (2025)

    Zhang, Y., Hu, Y., Song, Y., Zou, D., Lin, W.: Learning vision-based agile flight via differentiable physics. Nature Machine Intelligence7(6), 954–966 (2025)

    2025

  21. [21]

    Nature Machine Intelligence3(3), 218–229 (2021)

    Lu, L., Jin, P., Pang, G., Zhang, Z., Karniadakis, G.E.: Learning nonlinear oper- ators via DeepONet based on the universal approximation theorem of operators. Nature Machine Intelligence3(3), 218–229 (2021)

    2021

  22. [22]

    Nature Machine Intelligence 8, 1173–1181 (2026)

    Berner, J., Liu-Schiaffini, M., Kossaifi, J., Duruisseaux, V., Bonev, B., Aziz- zadenesheli, K., Anandkumar, A.: Principled approaches for extending neural architectures to function spaces for operator learning. Nature Machine Intelligence 8, 1173–1181 (2026)

    2026

  23. [23]

    Science Advances 7(40), 8605 (2021)

    Wang, S., Wang, H., Perdikaris, P.: Learning the solution operator of parametric partial differential equations with physics-informed deeponets. Science Advances 7(40), 8605 (2021)

    2021

  24. [24]

    Proceedings of the National Academy of Sciences120(39), 2310142120 (2023)

    Yang, L., Liu, S., Meng, T., Osher, S.J.: In-context operator learning with data prompts for differential equation problems. Proceedings of the National Academy of Sciences120(39), 2310142120 (2023)

    2023

  25. [25]

    Quantum9, 1761 (2025)

    Xiao, P., Zheng, M., Jiao, A., Yang, X., Lu, L.: Quantum DeepONet: Neural operators accelerated by quantum computing. Quantum9, 1761 (2025)

    2025

  26. [26]

    Preprint at https://arxiv.org/abs/2512 .24897 (2025)

    Xu, W., Han, J., Lai, R.: Self-Supervised Amortized Neural Operators for Optimal Control: Scaling Laws and Applications. Preprint at https://arxiv.org/abs/2512 .24897 (2025)

    2025

  27. [27]

    Journal of Computational Physics537, 114057 (2025)

    Huang, H., Lai, R.: Unsupervised solution operator learning for mean-field games. Journal of Computational Physics537, 114057 (2025)

    2025

  28. [28]

    Preprint at https://arxiv.org/abs/2601.09979 (2026)

    Cole, F., Wang, D., Chen, Y., Lu, Y., Lai, R.: In-Context Operator Learning on the Space of Probability Measures. Preprint at https://arxiv.org/abs/2601.09979 (2026)

    arXiv 2026

  29. [29]

    Mathematical Programming106(1), 25–57 (2006)

    W¨ achter, A., Biegler, L.T.: On the implementation of an interior-point fil- ter line-search algorithm for large-scale nonlinear programming. Mathematical Programming106(1), 25–57 (2006)

    2006

  30. [30]

    In: AIAA/AAS Astrodynamics Specialist Conference and Exhibit

    Vallado, D.A., Crawford, P., Hujsak, R., Kelso, T.S.: Revisiting spacetrack report #3. In: AIAA/AAS Astrodynamics Specialist Conference and Exhibit. American Institute of Aeronautics and Astronautics, Keystone, CO (2006). 27

    2006