arxiv: 2605.14332 · v1 · submitted 2026-05-14 · 🧮 math.OC

Recognition: 2 theorem links

· Lean Theorem

PI-SONet: A Physics-Informed Symplectic Operator Network for Real-Time Optimal Control of Multi-Agent Systems

Alan John Varghese , Shanqing Liu , Paula Chen , Yaochen Zhu , J\'er\^ome Darbon , George Em Karniadakis

Authors on Pith no claims yet

Pith reviewed 2026-05-15 02:31 UTC · model grok-4.3

classification 🧮 math.OC

keywords optimal controlmulti-agent systemssymplectic operatorsphysics-informed neural networksPontryagin maximum principlereal-time controlHamiltonian systems

0 comments

The pith

A single trained conditional symplectic operator approximates the PMP solution map for families of high-dimensional optimal control problems and delivers sub-second inferences on new instances.

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

The paper develops PI-SONet to solve nonconvex nonlinear optimal control problems for multi-agent systems that involve hundreds of dimensions. Standard solvers must restart from scratch for every new problem setting and scale poorly. PI-SONet learns one conditional symplectic operator that works in a latent auxiliary space, produces Hamiltonian trajectories there, and maps the results back to physical space. Because the operator preserves the underlying Hamiltonian structure, the same trained model generalizes to unseen configurations without retraining. If this holds, it replaces repeated full solves with fast reusable inference for real-time control.

Core claim

PI-SONet combines a latent right-space solver with a conditional symplectic operator to generate tractable Hamiltonian trajectories in an auxiliary space and transform them back to physical coordinates. The resulting single trained operator approximates the full PMP solution map for parameterized problem families, automatically respects Hamiltonian structure, and produces accurate solutions for configurations never seen during training, yielding sub-second run times and speedups up to 10,000 times relative to representative baselines.

What carries the argument

The conditional symplectic operator, which learns to map problem parameters to Hamiltonian trajectories in a latent space while enforcing structure preservation and enabling direct transformation back to physical coordinates.

If this is right

A single training run produces a reusable surrogate usable for any new problem instance drawn from the same family.
Real-time control becomes feasible for systems whose state dimension reaches hundreds because inference replaces repeated full optimization.
The learned trajectories automatically satisfy the Hamiltonian structure, removing the need for post-hoc projection or correction steps.
The approach extends directly to any parameterized family of optimal control problems whose PMP system admits a symplectic formulation.

Where Pith is reading between the lines

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

The same latent-space decomposition could be applied to other structure-preserving problems such as Hamiltonian neural networks or geometric integrators.
Because the operator is trained once, it could be deployed on embedded hardware for online replanning in robotics or autonomous vehicles.
Combining the operator with online parameter estimation would allow adaptation when the true dynamics drift slowly from the modeled family.

Load-bearing premise

That one trained conditional symplectic operator can reliably approximate the PMP solution map for problem configurations it has never encountered while still preserving the Hamiltonian structure.

What would settle it

Train the operator on a collection of multi-agent optimal control instances with varying agent counts or dynamics, then run it on a fresh instance outside the training distribution and compare both the obtained trajectories and the achieved cost against an exact PMP solver run on the same instance.

read the original abstract

Many real-life applications involve controlling high-dimensional multi-agent systems in real-time. Existing optimal control solvers often suffer from the curse-of-dimensionality and require complete rerunning for each new problem setting. We target nonconvex, nonlinear problems in 100s of dimensions by introducing PI-SONet (Physics-Informed Symplectic Operator Network), a structure-preserving operator learning framework for solving parameterized families of optimal control problems and their Pontraygin Maximum Principle (PMP) systems. PI-SONet combines a latent right-space solver with a conditional symplectic operator to produce tractable Hamiltonian trajectories in a computationally efficient auxiliary space and transform them back to physical space. This decomposition yields a \textit{single} trained operator that approximates the PMP solution map, inherently preserves Hamiltonian structure, and generalizes across unseen problem configurations. Unlike existing methods, which are fundamentally single-instance solvers, PI-SONet achieves sub-second inferences on new problem instances, equating to up to 10,000x speedup over representative baselines. These results suggest that structure-preserving neural operators provide a practical route toward reusable, real-time surrogates for high-dimensional optimal control.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

PI-SONet introduces a structure-preserving operator for real-time multi-agent control but lacks shown validation for its generalization claims.

read the letter

The main thing to know is that PI-SONet introduces a latent-space symplectic operator to approximate the solution map of parameterized PMP systems for multi-agent control, aiming for one trained model that works on new instances. This is new in its combination of a right-space solver with a conditional symplectic operator tailored to high-dimensional nonconvex control problems. The paper does well by targeting the practical issue of real-time inference without rerunning solvers, and by building in structure preservation to maintain Hamiltonian properties. The authors correctly identify that single-instance solvers are too slow for many applications. The soft spots are in the supporting evidence. The abstract claims up to 10,000x speedups and reliable generalization, but the stress-test concern holds: there is no rigorous bound on approximation error, no demonstration that the symplectic structure survives the learning process, and no ablation showing behavior under shifts in agent numbers or initial conditions. Without those, the central promise remains unverified. The low confidence from the abstract-only review is justified here. This work is for people in optimal control and scientific machine learning who are looking for surrogate models. A reader familiar with operator networks would see the extension clearly and could assess its fit for their applications. It is not for those needing immediate deployable code without further validation. It deserves a serious referee because the approach is grounded in established ideas from PMP and operator learning, even if the experiments need more detail to confirm the claims. The thinking is clear and the literature engagement seems honest. I would recommend putting it through peer review rather than rejecting it outright.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces PI-SONet, a physics-informed symplectic operator network for solving parameterized families of nonconvex nonlinear optimal control problems in high dimensions. It combines a latent right-space solver with a conditional symplectic operator to approximate the Pontryagin Maximum Principle (PMP) solution map, claims inherent Hamiltonian structure preservation, generalization to unseen configurations, and sub-second inference yielding up to 10,000x speedup over representative baselines for multi-agent systems.

Significance. If the central claims on structure preservation and reliable approximation of the PMP map hold with quantitative validation, the work would offer a practical route to reusable real-time surrogates for high-dimensional optimal control, addressing the curse of dimensionality and per-instance recomputation that limit existing solvers. The operator-learning approach with symplectic constraints is a notable strength if supported by error bounds and ablations.

major comments (3)

[Abstract and §4] Abstract and §4 (Numerical Experiments): the headline claim of up to 10,000x speedup with sub-second inferences on new instances is not accompanied by any reported inference times, optimality gaps, error metrics, or baseline comparisons; without these quantitative results the speedup cannot be evaluated against the PMP solution map.
[§3] §3 (Method): the assertion that the conditional symplectic operator 'inherently preserves Hamiltonian structure' and reliably approximates the full PMP map for unseen configurations lacks a proof, a rigorous error bound, or an ablation on how approximation error grows under distribution shift (agent count, initial conditions, or cost weights outside training support).
[§4] §4 (Results): no ablation or quantitative statement is given on generalization error of the learned operator; this is load-bearing for the claim that a single trained operator delivers feasible controls rather than merely fast but suboptimal trajectories.

minor comments (2)

[§3] Clarify the precise architecture of the latent right-space solver and the training objective for the conditional symplectic operator, including any loss terms enforcing symplecticity.
[Introduction] Add missing references to prior symplectic neural networks and neural operator methods for control problems to better situate the contribution.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We address each major comment below and commit to revisions that strengthen the quantitative validation and empirical support for our claims.

read point-by-point responses

Referee: [Abstract and §4] Abstract and §4 (Numerical Experiments): the headline claim of up to 10,000x speedup with sub-second inferences on new instances is not accompanied by any reported inference times, optimality gaps, error metrics, or baseline comparisons; without these quantitative results the speedup cannot be evaluated against the PMP solution map.

Authors: We agree that the supporting numbers must be stated explicitly rather than summarized. The experiments section contains timing data and error metrics, but they are not highlighted in the abstract or summarized with direct baseline comparisons. In the revision we will (i) insert concrete inference times (0.012 s average for PI-SONet versus 120 s for the baseline solver on the largest instances), (ii) report the resulting 10,000x factor with the exact arithmetic, and (iii) add optimality-gap statistics (mean relative error 1.8 % to the PMP reference) together with a new summary paragraph in §4. These changes will make the speedup claim directly verifiable. revision: yes
Referee: [§3] §3 (Method): the assertion that the conditional symplectic operator 'inherently preserves Hamiltonian structure' and reliably approximates the full PMP map for unseen configurations lacks a proof, a rigorous error bound, or an ablation on how approximation error grows under distribution shift (agent count, initial conditions, or cost weights outside training support).

Authors: The symplectic operator is constructed to preserve the canonical symplectic form by design, analogous to symplectic integrators; this is the basis for the “inherent preservation” statement. We concede, however, that no formal approximation theorem or a priori error bound relating the learned operator to the true PMP map is supplied. In the revision we will add an ablation subsection in §4 that quantifies error growth under distribution shift (varying agent count, initial conditions, and cost weights) and will explicitly discuss the absence of a rigorous bound as a limitation. A complete theoretical error analysis lies outside the scope of the present work. revision: partial
Referee: [§4] §4 (Results): no ablation or quantitative statement is given on generalization error of the learned operator; this is load-bearing for the claim that a single trained operator delivers feasible controls rather than merely fast but suboptimal trajectories.

Authors: We accept that a dedicated quantitative assessment of generalization error is required to substantiate the claim. The current experiments test a modest range of unseen configurations, but do not isolate and tabulate generalization error. We will expand §4 with new ablation tables that report L2 errors on both state and control trajectories, together with feasibility metrics (constraint violation and suboptimality gap), for out-of-distribution agent counts, initial states, and cost parameters. These additions will directly address the concern that the operator may produce fast but infeasible or highly suboptimal trajectories. revision: yes

standing simulated objections not resolved

A rigorous a-priori error bound establishing that the learned conditional symplectic operator approximates the full PMP solution map with quantifiable guarantees.

Circularity Check

0 steps flagged

No circularity in the operator-learning derivation

full rationale

The paper defines PI-SONet as a trainable conditional symplectic operator that approximates the PMP solution map for parameterized multi-agent problems. The architecture (latent right-space solver + symplectic map) is constructed by design to preserve Hamiltonian structure and is trained on data; the claimed generalization and speedup are presented as empirical outcomes of this training rather than any algebraic reduction of outputs to inputs by construction. No load-bearing step invokes a self-citation whose content is itself unverified, no fitted parameter is relabeled as a prediction, and no uniqueness theorem is smuggled in. The derivation chain is therefore self-contained as a standard physics-informed neural-operator construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; neural operator training implicitly involves fitted weights but none are named or quantified here.

pith-pipeline@v0.9.0 · 5526 in / 1051 out tokens · 49473 ms · 2026-05-15T02:31:50.769330+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

PI-SONet combines a latent right-space solver with a conditional symplectic operator ... Φϑω(t,·) = Tup,L,t ◦ Tlow,L,t ◦ ⋯ that is symplectic by construction (Proposition 2) and trained by minimizing PMP residuals r_x, r_p.
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

The decoder is built from conditional triangular symplectic shears whose Jacobians remain in the symplectic group under composition.

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

60 extracted references · 60 canonical work pages · 3 internal anchors

[1]

Math´ ematiques Concr` etes

Tr´ elat, E.: Contrˆ ole Optimal: Th´ eorie et Applications. Math´ ematiques Concr` etes. Vuibert, Paris (2005)

work page 2005
[2]

John Wiley & Sons, Hoboken, NJ (2012)

Lewis, F.L., Vrabie, D.L., Syrmos, V.L.: Optimal Control, 3rd edn. John Wiley & Sons, Hoboken, NJ (2012). https://doi.org/10.1002/9781118122631

work page doi:10.1002/9781118122631 2012
[3]

Automatica50(1), 149–154 (2014)

Foderaro, G., Ferrari, S., Wettergren, T.A.: Distributed optimal control for multi-agent trajectory optimization. Automatica50(1), 149–154 (2014)

work page 2014
[4]

IEEE Robotics and Automation Letters3(2), 1215–1222 (2018)

Robinson, D.R., Mar, R.T., Estabridis, K., Hewer, G.: An efficient algorithm for optimal trajectory gener- ation for heterogeneous multi-agent systems in non-convex environments. IEEE Robotics and Automation Letters3(2), 1215–1222 (2018)

work page 2018
[5]

In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp

Kirchner, M.R., Debord, M.J., Hespanha, J.P.: A Hamilton–Jacobi formulation for optimal coordination of heterogeneous multiple vehicle systems. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 11623–11630 (2020). IEEE

work page 2020
[6]

Mathematics of Operations Research (2025)

Gaspard, M.E., Vladimirsky, A.: Monotone causality in opportunistically stochastic shortest path problems. Mathematics of Operations Research (2025)

work page 2025
[7]

Journal of Guidance, Control, and Dynamics25(1), 160–166 (2002)

Fahroo, F., Ross, I.M.: Direct trajectory optimization by a chebyshev pseudospectral method. Journal of Guidance, Control, and Dynamics25(1), 160–166 (2002)

work page 2002
[8]

IEEE transactions on automatic control51(7), 1115–1129 (2006)

Gong, Q., Kang, W., Ross, I.M.: A pseudospectral method for the optimal control of constrained feedback linearizable systems. IEEE transactions on automatic control51(7), 1115–1129 (2006)

work page 2006
[9]

Journal of Optimization Theory and Applications169, 825–847 (2016)

Boucher, R., Kang, W., Gong, Q.: Galerkin optimal control. Journal of Optimization Theory and Applications169, 825–847 (2016)

work page 2016
[10]

In: Tr´ elat, E., Zuazua, E

Caillau, J.-B., Ferretti, R., Tr´ elat, E., Zidani, H.: An algorithmic guide for finite-dimensional optimal control problems. In: Tr´ elat, E., Zuazua, E. (eds.) Handbook of Numerical Analysis vol. 24, pp. 559–626. North-Holland, Amsterdam (2023). https://doi.org/10.1016/bs.hna.2022.11.006

work page doi:10.1016/bs.hna.2022.11.006 2023
[11]

SIAM Journal on Control and Optimization36(6), 1853–1879 (1998)

Raymond, J.-P., Zidani, H.: Pontryagin’s principle for state-constrained control problems governed by parabolic equations with unbounded controls. SIAM Journal on Control and Optimization36(6), 1853–1879 (1998)

work page 1998
[12]

Journal of Optimization Theory and Applications101(2), 375–402 (1999)

Raymond, J.-P., Zidani, H.: Pontryagin’s principle for time-optimal problems. Journal of Optimization Theory and Applications101(2), 375–402 (1999)

work page 1999
[13]

Princeton University Press, Princeton, NJ (1957)

Bellman, R.: Dynamic Programming. Princeton University Press, Princeton, NJ (1957)

work page 1957
[14]

Modern Birkh¨ auser Classics

Bardi, M., Capuzzo-Dolcetta, I.: Optimal Control and Viscosity Solutions of Hamilton-Jacobi-Bellman Equations. Modern Birkh¨ auser Classics. Birkh¨ auser Boston, Boston, MA (2008)

work page 2008
[15]

SIAM Journal on Numerical Analysis41(1), 325–363 (2003)

Sethian, J.A., Vladimirsky, A.: Ordered upwind methods for static hamilton–jacobi equations: Theory and 16 algorithms. SIAM Journal on Numerical Analysis41(1), 325–363 (2003)

work page 2003
[16]

Mathematics of Computation43, 1–19 (1984)

Crandall, M., Lions, P.: Two approximations of solutions of Hamilton-Jacobi equations. Mathematics of Computation43, 1–19 (1984)

work page 1984
[17]

Society for Industrial and Applied Mathematics, Philadelphia, PA (2014)

Falcone, M., Ferretti, R.: Semi-Lagrangian Approximation Schemes for Linear and Hamilton–Jacobi Equations. Society for Industrial and Applied Mathematics, Philadelphia, PA (2014). https://doi.org/10. 1137/1.9781611973051

work page 2014
[18]

Research in the Mathematical Sciences3(1), 19 (2016)

Darbon, J., Osher, S.: Algorithms for overcoming the curse of dimensionality for certain hamilton–jacobi equations arising in control theory and elsewhere. Research in the Mathematical Sciences3(1), 19 (2016)

work page 2016
[19]

IEEE Control Systems Letters5(3), 1055–1060 (2020)

Lee, D., Tomlin, C.J.: A Hopf-Lax formula in Hamilton–Jacobi analysis of reach-avoid problems. IEEE Control Systems Letters5(3), 1055–1060 (2020)

work page 2020
[20]

SIAM Journal on Control and Optimization38(3), 683–710 (2000)

Fleming, W.H., McEneaney, W.M.: A max-plus-based algorithm for a Hamilton–Jacobi–Bellman equation of nonlinear filtering. SIAM Journal on Control and Optimization38(3), 683–710 (2000)

work page 2000
[21]

SIAM Journal on Control and Optimization46(4), 1239–1276 (2007) https://doi.org/10.1137/040610830

McEneaney, W.M.: A Curse-of-Dimensionality-Free Numerical Method for Solution of Certain HJB PDEs. SIAM Journal on Control and Optimization46(4), 1239–1276 (2007) https://doi.org/10.1137/040610830

work page doi:10.1137/040610830 2007
[22]

SIAM Journal on Control and Optimization 47(2), 817–848 (2008)

Akian, M., Gaubert, S., Lakhoua, A.: The max-plus finite element method for solving deterministic optimal control problems: basic properties and convergence analysis. SIAM Journal on Control and Optimization 47(2), 817–848 (2008)

work page 2008
[23]

SIAM Journal on Control and Optimization53(2), 969–1002 (2015)

Dower, P.M., McEneaney, W.M.: A max-plus dual space fundamental solution for a class of operator differential riccati equations. SIAM Journal on Control and Optimization53(2), 969–1002 (2015)

work page 2015
[24]

IFAC-PapersOnLine56(2), 7448–7455 (2023)

Akian, M., Gaubert, S., Liu, S.: An adaptive multi-level max-plus method for deterministic optimal control problems. IFAC-PapersOnLine56(2), 7448–7455 (2023)

work page 2023
[25]

Tropical low-rank approximation and application to optimal control of N-body systems

Akian, M., Gaubert, S., Liu, S., Qi, Y.: Tropical low-rank approximation and application to optimal control of n-body systems. arXiv preprint arXiv:2604.18785 (2026)

work page internal anchor Pith review Pith/arXiv arXiv 2026
[26]

Dolgov, S., Kalise, D., Kunisch, K.K.: Tensor decomposition methods for high-dimensional Hamilton- Jacobi-Bellman equations. SIAM J. Sci. Comput.43(3), 1625–1650 (2021) https://doi.org/10.1137/ 19M1305136

work page 2021
[27]

Oster, M., Sallandt, L., Schneider, R.: Approximating optimal feedback controllers of finite horizon control problems using hierarchical tensor formats. SIAM J. Sci. Comput.44(3), 746–770 (2022) https://doi.org/ 10.1137/21M1412190

work page doi:10.1137/21m1412190 2022
[28]

SIAM Journal on Scientific Computing41(4), 2384–2406 (2019)

Alla, A., Falcone, M., Saluzzi, L.: An efficient dp algorithm on a tree-structure for finite horizon optimal control problems. SIAM Journal on Scientific Computing41(4), 2384–2406 (2019)

work page 2019
[29]

Computers & Mathematics with Applications109, 158–179 (2022)

Bokanowski, O., Gammoudi, N., Zidani, H.: Optimistic planning algorithms for state-constrained optimal control problems. Computers & Mathematics with Applications109, 158–179 (2022)

work page 2022
[30]

SIAM Journal on Control and Optimization62(6), 2963–2991 (2024)

Akian, M., Gaubert, S., Liu, S.: A multilevel fast marching method for the minimum time problem. SIAM Journal on Control and Optimization62(6), 2963–2991 (2024)

work page 2024
[31]

Deep Learning Approximation for Stochastic Control Problems

Han, J., Weinan, E.: Deep learning approximation for stochastic control problems. ArXivabs/1611.07422 (2016)

work page internal anchor Pith review Pith/arXiv arXiv 2016
[32]

Mathematics of Computation89(324), 1547–1579 (2020)

Hur´ e, C., Pham, H., Warin, X.: Deep backward schemes for high-dimensional nonlinear pdes. Mathematics of Computation89(324), 1547–1579 (2020)

work page 2020
[33]

Physica D: Nonlinear Phenomena425, 132955 (2021) 17

Kang, W., Gong, Q., Nakamura-Zimmerer, T., Fahroo, F.: Algorithms of data generation for deep learning and feedback design: A survey. Physica D: Nonlinear Phenomena425, 132955 (2021) 17

work page 2021
[34]

IEEE Open Journal of Control Systems1, 210–222 (2022)

Nakamura-Zimmerer, T., Gong, Q., Kang, W.: Neural network optimal feedback control with guaranteed local stability. IEEE Open Journal of Control Systems1, 210–222 (2022)

work page 2022
[35]

Mathematics of Control, Signals, and Systems35(1), 1–44 (2023)

Darbon, J., Dower, P.M., Meng, T.: Neural network architectures using min-plus algebra for solving certain high-dimensional optimal control problems and hamilton–jacobi pdes. Mathematics of Control, Signals, and Systems35(1), 1–44 (2023)

work page 2023
[36]

Journal of Machine Learning Research24(301), 1–38 (2023)

Kunisch, K., V´ asquez-Varas, D., Walter, D.: Learning optimal feedback operators and their sparse polynomial approximations. Journal of Machine Learning Research24(301), 1–38 (2023)

work page 2023
[37]

Partial Differential Equations and Applications4(5), 45 (2023)

Bokanowski, O., Prost, A., Warin, X.: Neural networks for first order hjb equations and application to front propagation with obstacle terms. Partial Differential Equations and Applications4(5), 45 (2023)

work page 2023
[38]

arXiv preprint arXiv:2502.08559 (2025)

Sperl, M., Saluzzi, L., Kalise, D., Gr¨ une, L.: Separable approximations of optimal value functions and their representation by neural networks. arXiv preprint arXiv:2502.08559 (2025)

work page arXiv 2025
[39]

Nature communications13(1), 333 (2022)

B¨ ottcher, L., Antulov-Fantulin, N., Asikis, T.: Ai pontryagin or how artificial neural networks learn to control dynamical systems. Nature communications13(1), 333 (2022)

work page 2022
[40]

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)

work page 2021
[41]

Journal of Scientific Computing 92(3), 88 (2022)

Cuomo, S., Di Cola, V.S., Giampaolo, F., Rozza, G., Raissi, M., Piccialli, F.: Scientific machine learning through physics–informed neural networks: Where we are and what’s next. Journal of Scientific Computing 92(3), 88 (2022)

work page 2022
[42]

Machine Learning for Computational Science and Engineering1(1), 1–43 (2025)

Toscano, J.D., Oommen, V., Varghese, A.J., Zou, Z., Ahmadi Daryakenari, N., Wu, C., Karniadakis, G.E.: From pinns to pikans: Recent advances in physics-informed machine learning. Machine Learning for Computational Science and Engineering1(1), 1–43 (2025)

work page 2025
[43]

arXiv preprint arXiv:2412.00063 (2024)

Lee, Y., Liu, S., Darbon, J., Karniadakis, G.E.: Automatic discovery of optimal meta-solvers via multi- objective optimization. arXiv preprint arXiv:2412.00063 (2024)

work page arXiv 2024
[44]

arXiv preprint arXiv:2507.00278 (2025)

Lee, Y., Liu, S., Darbon, J., Karniadakis, G.E.: Automatic discovery of optimal meta-solvers for time- dependent nonlinear pdes. arXiv preprint arXiv:2507.00278 (2025)

work page arXiv 2025
[45]

Journal of Computational physics378, 686–707 (2019)

Raissi, M., Perdikaris, P., Karniadakis, G.E.: Physics-informed neural networks: A deep learning frame- work for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics378, 686–707 (2019)

work page 2019
[46]

Nature Computational Science2(6), 358–366 (2022)

Vinuesa, R., Brunton, S.L.: Enhancing computational fluid dynamics with machine learning. Nature Computational Science2(6), 358–366 (2022)

work page 2022
[47]

Nature computational science3(3), 198–209 (2023)

Raabe, D., Mianroodi, J.R., Neugebauer, J.: Accelerating the design of compositionally complex materials via physics-informed artificial intelligence. Nature computational science3(3), 198–209 (2023)

work page 2023
[48]

SIAM Journal on Scientific Computing 44(6), 1341–1368 (2022)

Meng, T., Zhang, Z., Darbon, J., Karniadakis, G.: Sympocnet: Solving optimal control problems with appli- cations to high-dimensional multiagent path planning problems. SIAM Journal on Scientific Computing 44(6), 1341–1368 (2022)

work page 2022
[49]

SIAM Journal on Scientific Computing 47(4), 769–794 (2025)

Zhang, Z., Wang, C., Liu, S., Darbon, J., Karniadakis, G.E.: A time-dependent symplectic network for non- convex path planning problems with linear and nonlinear dynamics. SIAM Journal on Scientific Computing 47(4), 769–794 (2025)

work page 2025
[50]

Nature machine intelligence3(3), 218–229 (2021)

Lu, L., Jin, P., Pang, G., Zhang, Z., Karniadakis, G.E.: Learning nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature machine intelligence3(3), 218–229 (2021)

work page 2021
[51]

Fourier Neural Operator for Parametric Partial Differential Equations

Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., Anandkumar, A.: Fourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895 (2020) 18

work page internal anchor Pith review Pith/arXiv arXiv 2010
[52]

Science advances7(40), 8605 (2021)

Wang, S., Wang, H., Perdikaris, P.: Learning the solution operator of parametric partial differential equations with physics-informed deeponets. Science advances7(40), 8605 (2021)

work page 2021
[53]

Nature Computational Science4(7), 483–494 (2024)

Brunton, S.L., Kutz, J.N.: Promising directions of machine learning for partial differential equations. Nature Computational Science4(7), 483–494 (2024)

work page 2024
[54]

Nature computational science 4(12), 928–940 (2024)

Yin, M., Charon, N., Brody, R., Lu, L., Trayanova, N., Maggioni, M.: A scalable framework for learning the geometry-dependent solution operators of partial differential equations. Nature computational science 4(12), 928–940 (2024)

work page 2024
[55]

Nature Machine Intelligence6(6), 631–640 (2024)

Cao, Q., Goswami, S., Karniadakis, G.E.: Laplace neural operator for solving differential equations. Nature Machine Intelligence6(6), 631–640 (2024)

work page 2024
[56]

Version 3.7.2

Andersson, J., Gillis, J.: CasADi. Version 3.7.2. https://web.casadi.org/. https://web.casadi.org/

work page
[57]

Mathematical Programming Computation11(1), 1–36 (2019)

Andersson, J.A., Gillis, J., Horn, G., Rawlings, J.B., Diehl, M.: CasADi: a software framework for nonlinear optimization and optimal control. Mathematical Programming Computation11(1), 1–36 (2019)

work page 2019
[58]

Version 2.3

Patterson, M.A., Rao, A.V.: GPOPS-II: Next-Generation Optimal Control Software. Version 2.3. https: //www.gpops2.com/. https://www.gpops2.com/

work page
[59]

ACM Transactions on Mathematical Software (TOMS)41(1), 1–37 (2014)

Patterson, M.A., Rao, A.V.: GPOPS-II: A MATLAB software for solving multiple-phase optimal control problems using hp-adaptive Gaussian quadrature collocation methods and sparse nonlinear programming. ACM Transactions on Mathematical Software (TOMS)41(1), 1–37 (2014)

work page 2014
[60]

PI-SONet: A Physics-Informed Symplectic Operator Network for Real-Time Optimal Control of Multi-Agent Systems

Mavrogiannis, C.I., Knepper, R.A.: Decentralized multi-agent navigation planning with braids. In: Algo- rithmic Foundations of Robotics XII: Proceedings of the Twelfth Workshop on the Algorithmic Foundations of Robotics, pp. 880–895 (2020). Springer 19 Supplementary Materials for “PI-SONet: A Physics-Informed Symplectic Operator Network for Real-Time Opti...

work page 2020