Pith. sign in

REVIEW 4 major objections 5 minor 27 references

Physics-Constrained Generative Artificial Intelligence for Rapid Takeoff Trajectory Design

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Physics-GAN designs eVTOL takeoffs 200x faster and 99.6% accurate

desk verdict The physicsGAN idea is genuinely new and the speedup is impressive, but the headline feasibility claims are computed with the same surrogates used to train the model, so the feasible-space result is not yet shown to hold in the physics simulation. read the letter →

arxiv 2501.03445 v1 pith:2J5FYPAN submitted 2025-01-07 cs.LG

Samuel Sisk , Xiaosong Du This is my paper · ORCID
classification cs.LG
keywords physics-constrainedGANeVTOLtakeofftrajectorydesignsurrogate-basedoptimizationfeasiblespacegenerativeadversarialnetworkurbanairmobilityconstraintsatisfaction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes physicsGAN, a generative adversarial network that learns to output only feasible eVTOL takeoff control profiles. By conditioning on mass and power efficiency and penalizing the generator when surrogate-predicted acceleration exceeds a comfort limit, the model maps the original design space into a low-dimensional feasible space where roughly 98.9% of generated designs satisfy all five takeoff constraints. Combined with a surrogate-based differential evolution optimizer, the framework recovers the simulation-based reference optimum to 99.62% accuracy in 2.25 seconds, about 200 times faster than the SNOPT benchmark. The authors claim this is the first physics-constrained generative AI enabled by surrogate models for trajectory design.

What carries the argument

The physicsGAN architecture: a constraint-generator network that outputs takeoff duration and three sigmoid noise variables fed into two twinGAN generators for power and wing-angle B-spline control points; a discriminator; and surrogate models (a DNN for scalar energy and LSTM networks for time-series states) that compute the acceleration penalty $\lambda = (a_{\max}-0.3)^2$ during training. The surrogates let the physics penalty be evaluated without running the expensive simulation, which is what makes physics-constrained training tractable.

What would settle it

Run a few thousand physicsGAN-generated control profiles through the full simulation (not the surrogates) and measure the fraction that satisfies all five constraints and the resulting energy; if the simulation-verified coverage is far below the reported 98.9% or the best energy deviates from 1891 Wh by more than a few percent, the feasibility and accuracy claims would be undermined.

Watch

Extended reading notes

Core claim

The central claim is that a GAN whose generator is penalized during training by surrogate-predicted constraint violations can transform a high-dimensional, constrained design space into a low-dimensional, unconstrained feasible space. Specifically, the physicsGAN takes mass and efficiency as labels plus random noise, outputs a takeoff duration and three noise variables that feed twinGAN generators, which produce 40 B-spline control points for power and wing angle. The generator's loss includes a penalty $\lambda = (a_{\max}-0.3)^2$ whenever the surrogate-predicted peak acceleration exceeds 0.3 g; penalizing this single path constraint is enough, in practice, for roughly 98.9% of generated profiles to satisfy all five constraints. On the Airbus A3 Vahana example, differential evolution over the three physicsGAN noise variables with a DNN energy surrogate and LSTM state surrogates yields 1891.02 Wh, matching the 1883.84 Wh simulation-based SNOPT optimum with 99.62% relative accuracy, while always staying feasible and taking 2.25 seconds versus 533.38 seconds.

Load-bearing premise

The surrogate models trained on 5,504 simulation samples predict the constraints accurately enough for out-of-sample physicsGAN-generated designs, and the acceleration-only penalty is sufficient to make the other four constraints automatically satisfied.

Editorial extensions

If this is right

  • The proposed framework achieves 99.62% relative accuracy to the simulation-based optimal design in 2.25 seconds, a roughly 200x speedup over SNOPT.
  • Generated designs from physicsGAN satisfy all five takeoff constraints in about 98.9% of cases, so downstream optimization needs no explicit constraint handling.
  • Data-driven twinGAN with a gradient-based optimizer can get stuck in infeasible regions across random seeds, whereas physicsGAN remained feasible.
  • The design space reduction to three noise variables plus unconstrained optimization removes the cost of constraint evaluation entirely.
  • This is the first demonstration of a physics-constrained GAN enabled by surrogate models for engineering trajectory design, according to the authors' literature review.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the constraint coverage was computed with the same surrogates used for training, a natural next step is to verify coverage with high-fidelity simulation; if the surrogate errors are biased, the true coverage could be lower than 98.9%.
  • The acceleration-only penalty suggests a general recipe: enforce the binding constraint in the generator loss and rely on the learned data manifold to keep the other constraints satisfied; this may transfer to other trajectory design problems with similar constraint coupling.
  • The speedup depends on the simulation cost and the optimizer; on cheaper simulations or with many constraints, the feasibility advantage of physicsGAN may shrink relative to standard methods.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes physicsGAN, a GAN-based parameterization that maps a low-dimensional noise space to power and wing-angle control profiles for eVTOL takeoff trajectories, with the aim of generating only designs that satisfy all five takeoff constraints. The generator is trained with an additional penalty based on surrogate-predicted acceleration violations, and the resulting latent space is then used for unconstrained differential-evolution optimization. On the Airbus A3 Vahana example, the authors report 98.85% constraint coverage for the three-noise-variable physicsGAN, a 99.62% match to the SNOPT reference optimum at one design point, and a roughly 200x speedup. The paper also compares against a data-driven twinGAN baseline with derivative-free and gradient-based optimizers.

Significance. If the central claim is valid, the contribution is practically useful: it reduces a 41-dimensional constrained optimization problem to a 3-dimensional unconstrained one, removes the need to evaluate constraints during optimization, and demonstrates order-of-magnitude speedups on a credible eVTOL takeoff model. The comparison against two optimizer variants of the data-driven twinGAN is informative and gives the paper a clear empirical grounding. The idea of using surrogate-computed penalties to shape a generative model's latent space is transferable beyond this specific application. However, the paper's feasibility claims are currently measured through the same surrogates used to enforce them, so the central result needs external validation before the claimed feasible-space transformation can be accepted. The paper does not provide code or data, but the problem formulation is clear enough that the missing validation could in principle be supplied by the authors.

major comments (4)
  1. [Sec. III.C, Table 3] The coverage percentage is computed by checking constraints with the same surrogate models used to compute the physicsGAN penalty (Sec. II.D) and to define the feasible training set. This makes the 98.85% coverage a self-consistency metric rather than a measure of feasibility in the physical simulation. No simulation-based constraint check of generated physicsGAN designs is reported, so the central claim that physicsGAN transforms the design space into a feasible space is not externally validated. Please add a quantitative verification of generated trajectories against the original sim to confirm the surrogate predictions.
  2. [Sec. II.D and Sec. IV] The generator penalty is based only on the acceleration constraint, lambda = (a_max - 0.3)^2 for a_max >= 0.3, while the problem formulation in Table 1 has five constraints. The paper's own conclusion states that physicsGAN was 'only tested on the acceleration constraint.' There is no argument or evidence that satisfying the acceleration bound implies satisfaction of the final displacement, final velocity, and nonnegative-altitude constraints, yet Table 3 and the abstract report coverage against 'all constraints.' This gap must be closed, either by adding the other constraints to the training penalty or by demonstrating empirically that the acceleration-only penalty is sufficient on the simulation-verified set.
  3. [Sec. III.C, Table 4] The headline accuracy of 99.62% is obtained at a single point in the design requirement space (eta = 0.893, m = 723.85 kg). This does not establish that the framework delivers comparable accuracy across the mass and efficiency ranges used to train the surrogates and the GAN. Please report results at multiple design requirement points, ideally with a small grid or random sample, to support the claimed general capability of the proposed framework.
  4. [Abstract and Sec. III.C] The abstract states that the proposed framework achieves '100% feasible' results, but Table 3 reports 98.85% coverage for the three-noise-variable physicsGAN and Table 4 shows only that the single optimized result is marked 'Satisfied' by the surrogate-based check. The paper should distinguish between the coverage of the generated space and the feasibility of the optimized point, and should avoid claiming 100% feasibility unless it is supported by simulation-verified results across the space.
minor comments (5)
  1. [Sec. II.C] The description of the LSTM surrogates does not specify the input sequence length, the time discretization, or how the 500 output variables map to the five time-series responses (x, y, v_x, v_y, a). This information is needed for reproducibility.
  2. [Figs. 6 and 7] The legend entries 'TwinGAN Verification' and 'Constrained GAN Verification' are not defined in the text. Please state explicitly whether these curves come from running the original simulation model or from the surrogates, and if they are simulation-based, provide the corresponding quantitative errors.
  3. [Sec. IV] There is a grammatical error: 'the proposed physicsGAN were also compared' should be 'the proposed physicsGAN was also compared.' The same paragraph contains a sentence fragment ('Operating on the feasible space enables rapid, robust, accurate trajectory optimization') that should be attached to the preceding sentence or revised.
  4. [Table 4, twinGAN Trust-Constr rows] The 'Constraint Violation' column for the twinGAN Trust-Constr results lists percentages such as 31.96% and 0.05%. It is unclear whether these are relative violations of the worst constraint, the fraction of time steps violating a path constraint, or something else. Please define the metric.
  5. [Abstract and Sec. I] The claim that this is 'the first physics-constrained generative artificial intelligence enabled by surrogate models' is difficult to verify from the cited literature, which already includes constrained adversarial networks [23] and physics-informed GANs [20]. Please temper the novelty claim or provide a more precise distinction from these prior methods.

Circularity Check

2 steps flagged · score 6.0 of 10

Feasibility is defined, enforced, and measured by the same surrogates, so the reported 98.9% feasible-space coverage is a self-confirming metric rather than an external benchmark.

  1. fitted input called prediction [Sec. II.D (physicsGAN training) and Sec. III.C (coverage verification)]
    "Training the physicsGAN model requires a feasible dataset where all trajectories are feasible. This training dataset is created using random samples generated by the twinGAN model and verified by surrogates for constraint feasibility. ... The physicsGAN model with the highest coverage percentage during training is saved for surrogate-based optimization. ... It is calculated by randomly sampling the design requirements (i.e., 𝜂 and 𝑚) and design variables (i.e., physicsGAN noise variables) and checking the percentage of samples that meet the constraints."

    The same surrogate models are the sole arbiter of feasibility at every stage: they filter twinGAN samples into the 'feasible' training set, they compute the acceleration penalty λ=(a_max−0.3)^2 used to train the generator (Sec. II.D), they select the best physicsGAN checkpoint by coverage percentage, and they evaluate the final coverage in Table 3. The 98.85–99.79% coverage is therefore a measure of fit to the surrogate's constraint predictions, not a measure of physical constraint satisfaction. Because no simulation-based re-evaluation of the generated trajectories is reported, the 'transformed feasible space' claim is self-confirming; the held-out surrogate test accuracy (Table 2) only shows the surrogates approximate simulation on the training distribution, not on the generated designs.

  2. other [Sec. IV (Conclusions)]
    "The proposed physicsGAN-enabled surrogate-based optimization in this work was only tested on the acceleration constraint. A physicsGAN with more constraints will be investigated in the future work to provide a better comparison."

    This limitation directly contradicts the abstract's claim of 'around 98.9% of designs satisfying all constraints' and the five-constraint problem formulation in Sec. III.A. Only the acceleration constraint entered the training penalty; the other four constraints enter only through the surrogate-filtered feasible training set and the surrogate-based coverage check. The 'all constraints' feasible-space claim is therefore not independently verified in simulation, and the reported multi-constraint coverage is a surrogate self-assessment rather than a physical benchmark.

full rationale

The main derivation chain is: twinGAN parameterizes trajectories; surrogate models predict the states; physicsGAN is trained on a surrogate-filtered 'feasible' dataset with a surrogate-based acceleration penalty; the resulting feasible-space coverage is then assessed by checking constraints with the same surrogates. This loop makes the headline 98.85–99.79% coverage a self-confirming metric: the model is fit to the surrogate's notion of feasibility and then evaluated on that same notion. The surrogates themselves are trained and tested on simulation data (Sec. III.B), so the loop is not fully tautological, and the optimization accuracy comparison (99.62% vs SNOPT) has independent content. However, the paper reports no simulation re-evaluation of physicsGAN-generated trajectories, and the Conclusion admits that the physicsGAN was 'only tested on the acceleration constraint,' directly contradicting the abstract's 'satisfying all constraints' claim. The other four constraints enter only through the surrogate-filtered training set and surrogate-based coverage check, so the central 'feasible space' claim is partially circular and materially overstated. Score 6.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central feasibility claim rests on the simulation model, the surrogate models, and the assumption that an acceleration-only penalty yields a fully feasible space. No new physical entities are introduced.

free parameters (4)
  • Surrogate model weights (DNN for energy, LSTM for time-series states) = Trained on 5,504 samples; test accuracy over 99%
    The physicsGAN penalty and coverage metrics depend entirely on these surrogate predictions; their accuracy bounds the feasibility claim.
  • twinGAN generator weights = Trained on 1,099 optimal trajectories
    The physicsGAN reuses twinGAN generators to produce power and wing-angle control points; the feasible space is a subset of twinGAN's realistic space.
  • physicsGAN constraint-generator and discriminator weights = Trained on 10,601 feasible samples
    These weights define the mapping from (eta, m, z) to (t, z_t) that is claimed to produce feasible designs.
  • Latent noise dimension = 3
    The paper finds three noise variables give the best accuracy (99.62%) and acceptable coverage (98.85%); this dimensionality is a hand-chosen model hyperparameter.
assumptions (5)
  • domain assumption The Chauhan-Martins simulation model accurately represents the A3 Vahana takeoff dynamics.
    All training data and the reference optimal design come from this simulation; the paper does not validate against flight test data.
  • domain assumption Surrogate predictions generalize to physicsGAN-generated designs, including the optimized design.
    The optimized trajectory may lie outside the 5,504 training samples used for surrogates; the paper does not report surrogate error at the optimum.
  • ad hoc to paper Penalizing only the acceleration constraint yields a feasible space where displacement and velocity constraints are satisfied.
    The physicsGAN penalty uses lambda only for a_max; other constraints are expected to be satisfied through the feasible training dataset and adversarial training.
  • domain assumption The feasible training dataset (10,601 samples) is representative of the full design requirement space.
    Feasible samples are generated by filtering twinGAN random samples through surrogates; sampling density may not cover the boundary of the feasible region where the optimum lies.
  • domain assumption The twinGAN's 3-variable latent space sufficiently parameterizes optimal takeoff trajectories.
    The twinGAN was trained on 1,099 optimal trajectories and its reconstruction accuracy is around 99%.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics-Constrained Generative Artificial Intelligence for Rapid Takeoff Trajectory Design." pith.science (2026). https://pith.science/paper/2J5FYPAN

@misc{pith2026250103445,
  author       = {Pith},
  title        = {Pith review of: Physics-Constrained Generative Artificial Intelligence for Rapid Takeoff Trajectory Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2J5FYPAN}},
  note         = {Machine review of arXiv:2501.03445}
}
read the original abstract

To aid urban air mobility (UAM), electric vertical takeoff and landing (eVTOL) aircraft are being targeted. Conventional multidisciplinary analysis and optimization (MDAO) can be expensive, while surrogate-based optimization can struggle with challenging physical constraints. This work proposes physics-constrained generative adversarial networks (physicsGAN), to intelligently parameterize the takeoff control profiles of an eVTOL aircraft and to transform the original design space to a feasible space. Specifically, the transformed feasible space refers to a space where all designs directly satisfy all design constraints. The physicsGAN-enabled surrogate-based takeoff trajectory design framework was demonstrated on the Airbus A3 Vahana. The physicsGAN generated only feasible control profiles of power and wing angle in the feasible space with around 98.9% of designs satisfying all constraints. The proposed design framework obtained 99.6% accuracy compared with simulation-based optimal design and took only 2.2 seconds, which reduced the computational time by around 200 times. Meanwhile, data-driven GAN-enabled surrogate-based optimization took 21.9 seconds using a derivative-free optimizer, which was around an order of magnitude slower than the proposed framework. Moreover, the data-driven GAN-based optimization using gradient-based optimizers could not consistently find the optimal design during random trials and got stuck in an infeasible region, which is problematic in real practice. Therefore, the proposed physicsGAN-based design framework outperformed data-driven GAN-based design to the extent of efficiency (2.2 seconds), optimality (99.6% accurate), and feasibility (100% feasible). According to the literature review, this is the first physics-constrained generative artificial intelligence enabled by surrogate models.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 19 canonical work pages

  1. [1]

    Urban air mobility (UAM) market study,

    Goyal, R., Reiche, C., Fernando, C., Serrao, J., Kimmel, S., Cohen, A., and Shaheen, S., “Urban air mobility (UAM) market study,” Tech. rep., NASA, 2018

    work page 2018

  2. [2]

    Systems-levelanalysisofOnDemandMobilityforaviation,

    Vascik,P.D.,“Systems-levelanalysisofOnDemandMobilityforaviation,”Ph.D.thesis,MassachusettsInstituteofTechnology, 2017

    work page 2017

  3. [3]

    A study in reducing the cost of vertical flight with electric propulsion,

    Duffy, M. J., Wakayama, S. R., and Hupp, R., “A study in reducing the cost of vertical flight with electric propulsion,”17th AIAA Aviation Technology, Integration, and Operations Conference, 2017, p. 3442

    work page 2017

  4. [4]

    Challenges and key requirements of batteries for electric vertical takeoff and landing aircraft,

    Yang, X.-G., Liu, T., Ge, S., Rountree, E., and Wang, C.-Y., “Challenges and key requirements of batteries for electric vertical takeoff and landing aircraft,”Joule, Vol. 5, No. 7, 2021, pp. 1644–1659

    work page 2021

  5. [5]

    The promise of energy-efficient battery-powered urban aircraft,

    Sripad, S., and Viswanathan, V., “The promise of energy-efficient battery-powered urban aircraft,”Proceedings of the National Academy of Sciences, Vol. 118, No. 45, 2021, p. e2111164118

    work page 2021

  6. [6]

    Tilt-wing eVTOL takeoff trajectory optimization,

    Chauhan, S. S., and Martins, J. R., “Tilt-wing eVTOL takeoff trajectory optimization,”Journal of Aircraft, Vol. 57, No. 1, 2020, pp. 93–112

    2020

  7. [7]

    eVTOL passenger acceptance,

    Edwards, T., and Price, G., “eVTOL passenger acceptance,” Tech. rep., NASA, 2020

    work page 2020

  8. [8]

    Using multi-physics system simulation to predict battery pack thermal performance and risk of Thermal runaway during eVTOL aircraft operations,

    Harrison, J., Charles, D., Zenker, J., and Frank, E., “Using multi-physics system simulation to predict battery pack thermal performance and risk of Thermal runaway during eVTOL aircraft operations,”2019 AIAA/IEEE Electric Aircraft Technologies Symposium (EATS), IEEE, 2019, pp. 1–13. 9

    work page 2019

Show all 27 references
  1. [9]

    Simultaneous optimization of conceptual design and takeoff trajectory of a lift-plus-cruise UAV,

    Kaneko, S., and Martins, J., “Simultaneous optimization of conceptual design and takeoff trajectory of a lift-plus-cruise UAV,” 10th Autonomous VTOL Technical Meeting, 2023, pp. 1–18

    2023

  2. [10]

    Surrogate-based methods,

    Koziel, S., Ciaurri, D. E., and Leifsson, L., “Surrogate-based methods,”Computational optimization, methods and algorithms, 2011, pp. 33–59

    2011

  3. [11]

    Martins, J. R. R. A., and Ning, A.,Engineering Design Optimization, Cambridge University Press, Cambridge, UK, 2022. https://doi.org/10.1017/9781108980647, URL https://mdobook.github.io

    2022 doi

  4. [12]

    A new polynomial-time algorithm for linear programming,

    Karmarkar, N., “A new polynomial-time algorithm for linear programming,”Combinatorica, Vol. 4, No. 4, 1984, pp. 373–395. https://doi.org/10.1007/bf02579150

    1984 doi

  5. [13]

    SNOPT: An SQP algorithm for large-scale constrained optimization,

    Gill, P. E., Murray, W., and Saunders, M. A., “SNOPT: An SQP algorithm for large-scale constrained optimization,”SIAM Journal of Optimization, Vol. 12, No. 4, 2002, pp. 979–1006. https://doi.org/10.1137/S1052623499350013

    2002 doi

  6. [14]

    Machine learning in aerodynamic shape optimization,

    Li, J., Du, X., and Martins, J. R., “Machine learning in aerodynamic shape optimization,”Progress in Aerospace Sciences, Vol. 134, 2022, p. 100849

    2022

  7. [15]

    Survey of Multifidelity Methods in Uncertainty Propagation, Inference, and Optimization,

    Peherstorfer, B., Willcox, K., and Gunzburger, M., “Survey of Multifidelity Methods in Uncertainty Propagation, Inference, and Optimization,”SIAM Review, Vol. 60, No. 3, 2018, pp. 550–591. https://doi.org/10.1137/16M1082469

    2018 doi

  8. [16]

    Rapid airfoil design optimization via neural networks-based parameterization and surrogate modeling,

    Du, X., He, P., and Martins, J. R., “Rapid airfoil design optimization via neural networks-based parameterization and surrogate modeling,”Aerospace Science and Technology, Vol. 113, 2021, p. 106701

    2021

  9. [17]

    BezierGAN:AuomaticGenerationofSmoothCurvesfromInterpretableLow-DimensionalParameters,

    Chen,W.,andFuge,M.,“BezierGAN:AuomaticGenerationofSmoothCurvesfromInterpretableLow-DimensionalParameters,” arXiv:1808.08871, 2018

    2018 arXiv

  10. [18]

    Transfer-Learning-Enhanced Regression Generative Adversarial Networks for Optimal eVTOL Takeoff Trajectory Prediction,

    Yeh, S.-T., and Du, X., “Transfer-Learning-Enhanced Regression Generative Adversarial Networks for Optimal eVTOL Takeoff Trajectory Prediction,”Electronics, Vol. 13, No. 10, 2024, p. 1911

    2024

  11. [19]

    Generative adversarial networks (GAN) based efficient sampling of chemical composition space for inverse design of inorganic materials,

    Dan, Y., Zhao, Y., Li, X., Li, S., Hu, M., and Hu, J., “Generative adversarial networks (GAN) based efficient sampling of chemical composition space for inverse design of inorganic materials,”npj Computational Materials, Vol. 6, No. 1, 2020, p. 84

    2020

  12. [20]

    Physics-informed generative adversarial networks for stochastic differential equations,

    Yang, L., Zhang, D., and Karniadakis, G. E., “Physics-informed generative adversarial networks for stochastic differential equations,”SIAM Journal on Scientific Computing, Vol. 42, No. 1, 2020, pp. A292–A317

    2020

  13. [21]

    Physics-guided generative adversarial network for probabilistic structural system identification,

    Yu, Y., and Liu, Y., “Physics-guided generative adversarial network for probabilistic structural system identification,”Expert Systems with Applications, Vol. 239, 2024, p. 122339

    2024

  14. [22]

    Latent constraints: Learning to generate conditionally from unconditional generative models,

    Engel, J., Hoffman, M., and Roberts, A., “Latent constraints: Learning to generate conditionally from unconditional generative models,”arXiv preprint arXiv:1711.05772, 2017

    2017 arXiv

  15. [23]

    Efficient generation of structured objects with constrained adversarial networks,

    Di Liello, L., Ardino, P., Gobbi, J., Morettin, P., Teso, S., and Passerini, A., “Efficient generation of structured objects with constrained adversarial networks,”Advances in neural information processing systems, Vol. 33, 2020, pp. 14663–14674

    2020

  16. [24]

    Surrogate-Based Multidisciplinary Optimization for the Takeoff Trajectory Design of Electric Drones,

    Sisk, S., and Du, X., “Surrogate-Based Multidisciplinary Optimization for the Takeoff Trajectory Design of Electric Drones,” Processes, Vol. 12, No. 9, 2024, p. 1864

    2024

  17. [25]

    Long short-term memory,

    Hochreiter, S., and Schmidhuber, J., “Long short-term memory,”Neural computation, Vol. 9, No. 8, 1997, pp. 1735–1780

    1997

  18. [26]

    An efficient method for finding the minimum of a function of several variables without calculating derivatives,

    Powell, M. J., “An efficient method for finding the minimum of a function of several variables without calculating derivatives,” The computer journal, Vol. 7, No. 2, 1964, pp. 155–162

    1964

  19. [27]

    SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python,

    SciPy, “SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python,”Nature Methods, Vol. 17, 2020, pp. 261–272. https://doi.org/10.1038/s41592-019-0686-2. 10

    2020 doi

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.