Pith. sign in

REVIEW 4 major objections 6 minor 37 references

Physics-informed Gaussian Processes for Safe Envelope Expansion

T0 review · 4 major / 6 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 A physics-informed Gaussian process reads short-period modes from slow flight data.

desk verdict The paper's real claim—recovering short-period dynamics from near-steady rollercoaster data—is interesting but rests on an unproven kernel-scale assumption; the external historical comparison keeps it alive. read the letter →

arxiv 2501.01000 v1 pith:SM4XUG37 submitted 2025-01-02 cs.LG

D. Isaiah Harp , Joshua Ott , Dylan M. Asmar , John Alora , Mykel J. Kochenderfer This is my paper · ORCID
classification cs.LG
keywords Gaussianprocessesphysics-informedpriorsflighttestanalysisstabilityderivativesshort-periodmodepitchingmomentcoefficientenvelopeexpansionsystemidentification
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

The paper tries to establish that a Gaussian process whose mean function embeds an existing aerodynamic model can estimate the pitching moment coefficient from arbitrary, non-standard flight test data, and that the resulting surface, when differentiated, yields short-period frequency and damping estimates that match historical T-38 data. The authors deliberately start from a mismatched prior, the A-7E Morelli model, to show that the data can correct the prior rather than merely reproduce it. If true, this would remove the requirement for carefully repeated test points and allow envelope expansion from maneuvers pilots happen to fly, cutting the experimental campaign time.

What carries the argument

The load-bearing object is the Gaussian process posterior mean $\mu(\mathbf{x}) = m(\mathbf{x}) + K(\mathbf{x}, X)^\top A$, whose prior mean $m(\mathbf{x})$ is the polynomial aerodynamic prior drawn from a generic global aerodynamic model of a different aircraft (the A-7E), and whose kernel is the neural network kernel $k(\mathbf{x}, \mathbf{x}') = \sin^{-1}(\mathbf{x}^\top \mathbf{x}' / \sqrt{(1+\|\mathbf{x}\|^2/2)(1+\|\mathbf{x}'\|^2/2)})$. The argument proceeds by automatic differentiation of this posterior at trim conditions to get non-dimensional stability derivatives, conversion to dimensional derivatives via the standard longitudinal equations, and insertion into the short-period frequency and damping formulas (with $M_{\dot{\alpha}}$ replaced by $M_Q/3$, a standard textbook approximation). This differentiability is what turns the GP regression surface into a dynamical model.

What would settle it

A direct falsifier would be to take the same rollercoaster dataset, replace the polynomial aerodynamic prior with a zero mean, and re-derive $\omega_{SP}$ and $\zeta_{SP}$; if the short-period predictions remain accurate, the physics prior is not the cause of the frequency bridging, while if they degrade or become meaningless, the prior is doing the work.

Watch

Extended reading notes

Core claim

The central claim is that a physics-informed Gaussian process, using a polynomial aerodynamic model as its mean function and a neural network kernel, can turn low-frequency rollercoaster maneuver data (pilot input at about 0.0036 Hz, roughly 1% of the short-period frequency) into accurate predictions of the T-38C short-period frequency and damping. The GP posterior is differentiated at experimentally determined trim conditions to obtain stability derivatives $C_{m_\alpha}$ and $C_{m_Q}$, and an auxiliary Gaussian process on the Z-force coefficient yields $Z_\alpha$; these feed the standard linearized short-period formulas. Across Mach and dynamic pressure profiles, the continuous predictions of $\omega_{SP}$ and $\zeta_{SP}$ agree with consolidated historical data from four independent flight test sources, with root-mean-square errors comparable to the internal scatter of that historical data. The paper explicitly hypothesizes that the physics-based mean function is what lets high-frequency modes be recovered from low-frequency content.

Load-bearing premise

The whole argument rests on the assumption that a GP whose mean function is a deliberately mismatched aircraft model can bridge a roughly one-hundred-fold gap between the excitation frequency in the data and the short-period mode being identified; if the physics prior does all the work, the 'data-driven' short-period estimates are just artifacts of the prior.

Editorial extensions

If this is right

  • Flight test campaigns would no longer need predefined test points with tight tolerances; data from arbitrary maneuvers such as rollercoasters would suffice, potentially eliminating the roughly six-fold repetition overhead cited for a major envelope expansion program.
  • The GP can be queried at any physically achievable flight condition, producing continuous curves of short-period frequency and damping versus dynamic pressure instead of single-point estimates from doublets or sweeps.
  • Because the model updates an aerodynamic prior with flight data, it corrects a deliberately wrong prior toward the true aircraft's values, suggesting the same architecture can transfer models between airframes.
  • The approach extends beyond $C_m$: the same differentiation pipeline can estimate other stability derivatives, and the paper proposes multi-output Gaussian processes for a fuller characterization of aircraft dynamics.

Reading between the lines

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

  • If the frequency-bridging effect is real, it suggests the physics prior acts as a regularizer that anchors the GP's high-wavenumber behavior even when the data has no energy there; a testable extension is to remove the prior entirely and see whether the recovered short-period modes collapse.
  • The same reasoning might transfer to other under-determined dynamical identification problems, such as extracting structural modes from low-frequency ground vibrations or power-system oscillations from slow telemetry, whenever a credible physics-based prior exists.
  • A stronger validation of the bridging hypothesis would be to compare GP-derived $C_{m_\alpha}$ and $C_{m_Q}$ directly against frequency-sweep estimates at the same Mach and dynamic pressure, rather than only comparing the derived $\omega_{SP}$ and $\zeta_{SP}$.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper introduces a physics-informed Gaussian process approach to estimate the pitching moment coefficient C_m from arbitrary flight test data. A Morelli aerodynamic model (for the A-7E) is used as the mean function, and the GP maps a state vector of Mach, density, dynamic pressure, angular rates, angle of attack, and stabilator deflection to C_m. The authors compute stability derivatives C_m_alpha and C_m_Q by differentiating the posterior mean at trim conditions, then derive short-period frequency and damping, comparing with historical T-38 data. The central claim is that low-frequency (0.0036 Hz) rollercoaster maneuvers provide enough information to recover short-period dynamics across several Mach and dynamic pressure profiles.

Significance. If the central claim holds, the method could substantially reduce the flight test burden by removing the need for tightly toleranced test points, and the paper provides code and data. The use of external historical comparisons and the deliberate choice of a mismatched prior are honest and useful design features. However, the core mechanism is currently an explicit hypothesis: the paper does not demonstrate that the GP posterior actually refines the prior in the alpha and Q directions rather than inheriting the short-period content from the A-7 mean function. The significance therefore depends on the outcome of the tests requested below.

major comments (4)
  1. [Eqs. (4), (6), (11)-(13)] The neural-network kernel in Eq. (6) has no per-dimension lengthscale or automatic relevance determination, and the state vector in Eq. (4) uses raw physical units. With qbar ranging 100-800 lb/ft^2 and Mach 0.5-1.1 while alpha and Q are O(0.01-0.1), the inner product x^T x' is dominated by qbar and Mach. Consequently, for points differing mainly in alpha or Q the covariance k(x, X) is nearly constant in those coordinates, and the gradient term (∇_x K(x,X))^T A in Eq. (13) is negligible for the alpha and Q components. The posterior derivatives Cm_alpha and Cm_Q in Eqs. (9)-(10) then reduce to derivatives of the fixed A-7 Morelli prior, independent of the T-38 data. The manuscript never states that inputs are standardized or that the kernel incorporates dimension-specific scaling. As written, the central claim in Section V that the GP refines the prior toward the T-38 is unsupported. The authors should either standardize inputs (or use an ARD kernel) and demonstrate that the posterior Cm_alpha and Cm_Q differ materially from the prior values, or explicitly show that the data-correction term is non-negligible in the alpha and Q directions.
  2. [Section V, Fig. 2 and Table 2] The validation of Cm in Figure 2 is in-sample: the text does not state a train/test split, and the GP appears to be trained on the same rollercoaster maneuver used for the comparison. No baseline is provided against a GP with zero mean, a GP with the Morelli mean but no data (prior only), or a standard GP without the physics-informed mean. Consequently, the agreement in Figure 2 and the dynamic parameters in Table 2 cannot be attributed to the data-updating mechanism. The authors should add an out-of-sample evaluation (e.g., held-out maneuvers or time segments) and report ablations that isolate the contribution of the data versus the prior.
  3. [Section V, final paragraphs] The paper's central claim is explicitly conditional: 'We hypothesize that we were able to reproduce high-frequency modes with low-frequency experimental content due to our imposition of a physics-based prior.' This hypothesis is load-bearing, yet no evidence is provided to distinguish a genuinely data-driven refinement from a prior artifact. Specifically, the paper should report (a) the omega_SP and zeta_SP that would be predicted by the A-7 mean function alone at the same trim conditions, (b) the GP posterior predictions, and (c) the sensitivity of both to the kernel scaling. Without these, the agreement with the historical data in Figures 4-5 could be entirely explained by the prior mean, which would invalidate the claim that arbitrary low-frequency flight data suffice.
  4. [Section V, Table 3] The RMSE comparison in Table 3 is weakened by the unspecified thresholds for Mach regions (±0.02 Mach similarity) and dynamic pressure grouping (±40 lb/ft^2), and by the fact that some Mach regions have as few as two or three historical points. The 'Dataset' RMSE is therefore not a stable baseline, and the paper provides no uncertainty quantification on the RMSE values themselves. Please report the number of points in each group, the exact region boundaries, and bootstrap confidence intervals for the RMSE to support the claim that the predictions are 'relatively closely aligned' with the historical data.
minor comments (6)
  1. [Section V] The description of the rollercoaster excitation as 'AOA oscillations at a rate of around 1.3 deg/sec, or a mere 0.0036 Hz' is dimensionally unclear; state the oscillation frequency or period explicitly.
  2. [Eq. (15)] Provide a derivation of the damping ratio formula, including the sign conventions and the replacement of M_alpha_dot by (1/3)M_Q, and check the dimensional consistency of the expression.
  3. [Section IV.A.1] Report the fitted coefficients a, b, c, d for the trim regressions, the number of trim shots, and the goodness of fit, since the trim condition determines the differentiation point at which the stability derivatives are evaluated.
  4. [Section IV] State how the GP noise variance nu in Eq. (12) is determined: fixed a priori, optimized via maximum marginal likelihood, or chosen by cross-validation.
  5. [Section V, Fig. 2] Report the number of training points and maneuvers used to fit the GP, and clarify whether the data shown in Fig. 2 were part of the training set.
  6. [Fig. 3] Add a legend or explicit cross-reference to Table 1 for the symbols shown in Fig. 3, since the caption does not identify them.

Circularity Check

1 steps flagged · score 2.0 of 10

Central short-period validation is non-circular; only minor in-sample C_m plot is presented as prediction.

  1. fitted input called prediction [Section V, paragraph describing Figure 2]
    "These values are indicated by the dashed line. The solid line denotes the GP prediction and a shaded region captures the 95% confidence interval. The GP model successfully captures the general trend of C_m, providing reasonably accurate estimates while also quantifying the inherent uncertainty in dynamic maneuvers."

    The GP posterior mean in Eq. (11) is mu(x) = m(x) + K(x,X)^T A with A = (K(X,X) + nu I)^{-1}(y - m(X)), so at the training inputs the posterior is a shrinkage of the training observations y. The rollercoaster maneuver shown in Figure 2 is the same maneuver used to build the model, and the paper states no held-out split. Therefore the close agreement between the solid GP curve and the dashed 'true' C_m values is an in-sample fit assured by the regression construction, not evidence of predictive generalization. This in-sample display is not load-bearing for the short-period claim, because omega_SP and zeta_SP are compared to external historical data in Table 1 and Figures 3-5.

full rationale

The paper's central derivation—computing omega_SP and zeta_SP from GP derivatives at trim states and comparing them to external historical data (Table 1, Figures 3-5)—is self-contained against external benchmarks. The stability derivatives Cm_alpha and Cm_Q are differentiated from the GP posterior (Eq. 13) and inserted into standard linearized short-period formulas (Eqs. 14-15); neither formula is fit to the historical omega_SP or zeta_SP values. The Morelli mean function is an external, deliberately mismatched A-7 prior, not a self-citation, and the GP data-correction term is what is claimed to refine it. Self-citations such as [24], [25], and [27] are not load-bearing. The paper's own hypothesis that the physics-based prior may be responsible for reproducing high-frequency modes (Section V) and its admission that sparse historical data artificially lowers reported RMSE (Section V, Table 3) are limitations, not circular steps. The only in-sample presentation is Figure 2, where the GP is compared with the C_m values of the maneuver used for training, with no stated held-out split; that is a fitted input called a prediction, but it does not drive the central external validation. Overall score: 2.

Assumptions & free parameters 6 free parameters · 8 assumptions · 0 invented entities

The central claim rests on a set of domain assumptions about the validity of the intentionally mismatched aerodynamic prior, the short-period approximations, and the fitted trim regressions. The most fragile is the ad hoc assumption that a physics-informed GP can bridge a 100x frequency gap, which the paper itself labels a hypothesis. No new physical entities are introduced.

free parameters (6)
  • Trim regression coefficients a, b in alpha_trim(qbar) = Not reported
    Fit per Mach range from trim shots; they define the trim states at which the GP is differentiated (Eq. 7).
  • Trim regression coefficients c, d in delta_e_trim(qbar) = Not reported
    Fit per Mach range from trim shots; these define trim stabilator deflection for derivative evaluation (Eq. 8).
  • GP noise variance nu = Not reported
    Observation noise variance in Eqs. (1)-(2); the estimation procedure is not described.
  • Mach region boundaries for damping curves = 0.5, 0.7, 0.9, 1.08
    Hand-chosen thresholds for 'low', 'moderate', 'high', and 'supersonic' ranges used in Figure 5.
  • Dynamic pressure grouping interval for RMSE = +/-40 lb/ft^2
    Used to group historical data for the RMSE comparison in Table 3.
  • Mach similarity tolerance for damping comparison = +/-0.02
    Used to filter comparisons by Mach number for damping.
assumptions (8)
  • standard math Gaussian process regression equations (Eqs. 1-3) are correct
    The GP conditioning formulas are standard; no proof is needed.
  • domain assumption The Morelli model form with A-7E coefficients is a suitable prior for T-38C C_m
    The paper deliberately uses a prior known to be incorrect for the T-38C, yet assumes the functional form and coefficients provide useful inductive bias (Section IV).
  • domain assumption The neural network kernel (Eq. 6) is appropriate for this regression
    No justification beyond citing standard GP references is given for this choice of kernel.
  • domain assumption The 2-DOF short-period approximation and M_dot_alpha = M_Q/3 are valid for the T-38C
    Eqs. (14)-(15) rely on a textbook approximation with the replacement M_dot_alpha = 1/3 M_Q from Yechout [3]; no T-38-specific validation is provided.
  • domain assumption The trim regression forms for alpha_trim and delta_e_trim are accurate across Mach ranges
    Eqs. (7)-(8) choose exponential and logarithmic functional forms without a stated model-selection process; these fitted functions define the trim states for derivative evaluation.
  • domain assumption Historical comparison data are usable as ground truth despite unstandardized CG and standard atmosphere assumptions
    Section V acknowledges CG location alters short period frequency and some data lack dynamic pressure or actual temperature, yet still treats the data as valid benchmarks.
  • domain assumption The true C_m targets computed from Eq. (16) are accurate
    The moment equation relies on measured angular accelerations, products of inertia, and the assumption that the summation of moments is dominated by the listed terms.
  • ad hoc to paper Low-frequency (0.0036 Hz) excitation can yield high-frequency short-period dynamics through the GP with physics-informed prior
    This is the paper's central hypothesis, stated in Section V without derivation or dedicated supporting experiment; it is the load-bearing premise for the short-period predictions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics-informed Gaussian Processes for Safe Envelope Expansion." pith.science (2026). https://pith.science/paper/SM4XUG37

@misc{pith2026250101000,
  author       = {Pith},
  title        = {Pith review of: Physics-informed Gaussian Processes for Safe Envelope Expansion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SM4XUG37}},
  note         = {Machine review of arXiv:2501.01000}
}
read the original abstract

Flight test analysis often requires predefined test points with arbitrarily tight tolerances, leading to extensive and resource-intensive experimental campaigns. To address this challenge, we propose a novel approach to flight test analysis using Gaussian processes (GPs) with physics-informed mean functions to estimate aerodynamic quantities from arbitrary flight test data, validated using real T-38 aircraft data collected in collaboration with the United States Air Force Test Pilot School. We demonstrate our method by estimating the pitching moment coefficient without requiring predefined or repeated flight test points, significantly reducing the need for extensive experimental campaigns. Our approach incorporates aerodynamic models as priors within the GP framework, enhancing predictive accuracy across diverse flight conditions and providing robust uncertainty quantification. Key contributions include the integration of physics-based priors in a probabilistic model, which allows for precise computation from arbitrary flight test maneuvers, and the demonstration of our method capturing relevant dynamic characteristics such as short-period mode behavior. The proposed framework offers a scalable and generalizable solution for efficient data-driven flight test analysis and is able to accurately predict the short period frequency and damping for the T-38 across several Mach and dynamic pressure profiles.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 34 canonical work pages

  1. [1]

    Generic global aerodynamic model for aircraft,

    Grauer, J. A., and Morelli, E. A., “Generic global aerodynamic model for aircraft,”Journal of Aircraft, Vol. 52, No. 1, 2015, pp. 13–20

    work page 2015

  2. [2]

    IntroductiontotheAdvancesinAircraftSystemIdentificationfromFlightTestData,

    Grauer,J.A.,andMorelli,E.A.,“IntroductiontotheAdvancesinAircraftSystemIdentificationfromFlightTestData,” Journal of Aircraft, Vol. 60, No. 5, 2023, pp. 1329–1330

    work page 2023

  3. [3]

    R.,Introduction to Aircraft Flight Mechanics, AIAA, 2003

    Yechout, T. R.,Introduction to Aircraft Flight Mechanics, AIAA, 2003

    work page 2003

  4. [4]

    T-38A Category II Stability and Control Tests,

    Lusby, W. A., and Hanks, N. J., “T-38A Category II Stability and Control Tests,” Tech. Rep. AFFTC-TR-61-15, Air Force Flight Test Center, Edwards AFB, CA, Aug. 1961

    work page 1961

  5. [5]

    The New Era of Loads Flight Testing: Artificial Intelligence (AI) in Real-Time Envelope Expansion Testing,

    Luquin, V., McCoy, S. S., Garcia, O. R., Latkin, A. W., and Hashii, W. N., “The New Era of Loads Flight Testing: Artificial Intelligence (AI) in Real-Time Envelope Expansion Testing,”The Society of Flight Test Engineers Annual International Symposium, Vol. 53rd, 2023

    work page 2023

  6. [6]

    T-38C Limited Flight Test and Evaluation of Flying Qualities,

    TPS-23B, “T-38C Limited Flight Test and Evaluation of Flying Qualities,” Tech. Rep. TPS 23B, United States Air Force Test Pilot School, Edwards AFB, CA, Jan. 2024

    work page 2024

  7. [7]

    T-38 Peace Patch I Test Plan, Longitudinal and Lateral-Directional,

    TPS-23B, “T-38 Peace Patch I Test Plan, Longitudinal and Lateral-Directional,” Tech. Rep. TPS 23B, United States Air Force Test Pilot School, Edwards AFB, CA, Jan. 2024

    work page 2024

  8. [8]

    Global Nonlinear Aerodynamic Modeling Using Multivariate Orthogonal Functions,

    Morelli, E., “Global Nonlinear Aerodynamic Modeling Using Multivariate Orthogonal Functions,”Journal of Aircraft, Vol. 32, No. 2, 1995, p. 270–277. 16

    work page 1995

Show all 37 references
  1. [9]

    Ljung, L.,System Identification: Theory for the User, Pearson Education, 1998

    1998

  2. [10]

    L., and Kutz, J

    Brunton, S. L., and Kutz, J. N.,Data-driven science and engineering: Machine learning, dynamical systems, and control, Cambridge University Press, 2022

    2022

  3. [11]

    Limited aerodynamic System Identification of the T-38A using SIDPAC software,

    Shepherd, M. J., Jorris, T. R., and Gray, W. R., “Limited aerodynamic System Identification of the T-38A using SIDPAC software,”IEEE Aerospace Conference, 2010, pp. 1–10

    2010

  4. [12]

    T-38C Transfer Function Modeling in System Identification using Comprehensive Identification Frequency Response (CIFER),

    Welborn, C., “T-38C Transfer Function Modeling in System Identification using Comprehensive Identification Frequency Response (CIFER),” Master’s thesis, California Polytechnic State University, San Luis Obispo, CA, Aug. 2020

    2020

  5. [13]

    Tangirala, A.,Principles of System Identification: Theory and Practice, CRC Press, 2018

    2018

  6. [14]

    Keesman, K.,System Identification: An Introduction, Advanced Textbooks in Control and Signal Processing, Springer London, 2011

    2011

  7. [15]

    Problems in estimating dynamics from data,

    Kostelich, E. J., “Problems in estimating dynamics from data,”Physica D: Nonlinear Phenomena, Vol. 58, No. 1-4, 1992, pp. 138–152

    1992

  8. [16]

    Convex optimization techniques in system identification,

    Vandenberghe, L., “Convex optimization techniques in system identification,”IFAC Proceedings Volumes, Vol. 45, No. 16, 2012, pp. 71–76

    2012

  9. [17]

    Development of a Full Flight Envelope F-16 VISTA Simulation Model from Closed-loop Flight Data,

    Knapp, M., Berger, T., Tischler, M., Cotting, M. C., and Marcus, A., “Development of a Full Flight Envelope F-16 VISTA Simulation Model from Closed-loop Flight Data,”Proceedings of the AIAA Atmospheric Flight Mechanics Conference, AIAA SciTech Forum, 2018

    2018

  10. [18]

    Discovering governing equations from data by sparse identification of nonlinear dynamical systems,

    Brunton, S. L., Proctor, J. L., and Kutz, J. N., “Discovering governing equations from data by sparse identification of nonlinear dynamical systems,”Proceedings of the National Academy of Sciences, Vol. 113, No. 15, 2016, pp. 3932–3937

    2016

  11. [19]

    Sparse identification of nonlinear dynamics for model predictive control in the low-data limit,

    Kaiser, E., Kutz, J. N., and Brunton, S. L., “Sparse identification of nonlinear dynamics for model predictive control in the low-data limit,”Proceedings of the Royal Society, Vol. 474, No. 2219, 2018, p. 20180335

    2018

  12. [20]

    Input design for structured nonlinear system identification,

    Vincent, T. L., Novara, C., Hsu, K., and Poolla, K., “Input design for structured nonlinear system identification,”Automatica, Vol. 46, No. 6, 2010, pp. 990–998

    2010

  13. [21]

    Nonlinear system identification in structural dynamics: 10 more years of progress,

    Noël, J.-P., and Kerschen, G., “Nonlinear system identification in structural dynamics: 10 more years of progress,”Mechanical Systems and Signal Processing, Vol. 83, 2017, pp. 2–35

    2017

  14. [22]

    Learning to control linear systems can be hard,

    Tsiamis, A., Ziemann, I. M., Morari, M., Matni, N., and Pappas, G. J., “Learning to control linear systems can be hard,” Conference on Learning Theory, PMLR, 2022, pp. 3820–3857

    2022

  15. [23]

    Active Learning for Control-Oriented Identification of Nonlinear Systems,

    Lee, B. D., Ziemann, I., Pappas, G. J., and Matni, N., “Active Learning for Control-Oriented Identification of Nonlinear Systems,”arXiv preprint arXiv:2404.09030, 2024. 17

    2024 arXiv

  16. [24]

    Approximate sequential optimization for informative path planning,

    Ott, J., Kochenderfer, M. J., and Boyd, S., “Approximate sequential optimization for informative path planning,”Robotics and Autonomous Systems, Vol. 182, 2024, p. 104814

    2024

  17. [25]

    J., and Wheeler, T

    Kochenderfer, M. J., and Wheeler, T. A.,Algorithms for Optimization, MIT Press, 2019

    2019

  18. [26]

    K., and Rasmussen, C

    Williams, C. K., and Rasmussen, C. E.,Gaussian Processes for Machine Learning, MIT Press, 2006

    2006

  19. [27]

    Informative Input Design for Dynamic Mode Decomposition,

    Ott, J., Kochenderfer, M. J., and Boyd, S., “Informative Input Design for Dynamic Mode Decomposition,”arXiv preprint arXiv:2409.13088, 2024

    2024 arXiv

  20. [28]

    GPR-based novel approach for non-linear aerodynamic modelling from flight data,

    Kumar, A., and Ghosh, A. K., “GPR-based novel approach for non-linear aerodynamic modelling from flight data,”The Aeronautical Journal, Vol. 123, No. 1259, 2019, pp. 79–92

    2019

  21. [29]

    Physics-informed learning machines for partial differential equations: Gaussian processes versus neural networks,

    Pang, G., and Karniadakis, G. E., “Physics-informed learning machines for partial differential equations: Gaussian processes versus neural networks,”Emerging Frontiers in Nonlinear Science, 2020, pp. 323–343

    2020

  22. [30]

    Physics-informed CoKriging: A Gaussian-process- regression-based multifidelity method for data-model convergence,

    Yang, X., Barajas-Solano, D., Tartakovsky, G., and Tartakovsky, A. M., “Physics-informed CoKriging: A Gaussian-process- regression-based multifidelity method for data-model convergence,”Journal of Computational Physics, Vol. 395, 2019, pp. 410–431

    2019

  23. [31]

    Physics-informed Gaussian process regression for states estimation and forecasting in power grids,

    Tartakovsky, A. M., Ma, T., Barajas-Solano, D. A., and Tipireddy, R., “Physics-informed Gaussian process regression for states estimation and forecasting in power grids,”International Journal of Forecasting, Vol. 39, No. 2, 2023, pp. 967–980

    2023

  24. [32]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,

    Raissi, M., Perdikaris, P., and 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 physics, Vol. 378, 2019, pp. 686–707

    2019

  25. [33]

    Wind load estimation and virtual sensing in long-span suspension bridges using physics-informed Gaussian process latent force models,

    Petersen, Ø., Øiseth, O., and Lourens, E., “Wind load estimation and virtual sensing in long-span suspension bridges using physics-informed Gaussian process latent force models,”Mechanical Systems and Signal Processing, Vol. 170, 2022, p. 108742

    2022

  26. [34]

    A spectrum of physics-informed Gaussian processes for regression in engineering,

    Cross, E. J., Rogers, T. J., Pitchforth, D. J., Gibson, S. J., Zhang, S., and Jones, M. R., “A spectrum of physics-informed Gaussian processes for regression in engineering,”Data-Centric Engineering, Vol. 5, 2024, p. e8

    2024

  27. [35]

    Physics-informed machine learning,

    Karniadakis, G. E., Kevrekidis, I. G., Lu, L., Perdikaris, P., Wang, S., and Yang, L., “Physics-informed machine learning,” Nature Reviews Physics, Vol. 3, No. 6, 2021, pp. 422–440

    2021

  28. [36]

    Some new results on neural network approximation,

    Hornik, K., “Some new results on neural network approximation,”Neural Networks, Vol. 6, No. 8, 1993, pp. 1069–1072

    1993

  29. [37]

    C.,Flight Stability and Automatic Control, 2nd ed., McGraw-Hill, 1998

    Nelson, R. C.,Flight Stability and Automatic Control, 2nd ed., McGraw-Hill, 1998. 18

    1998

Pith tools

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