Pith. sign in

REVIEW 5 major objections 5 minor 55 references

Statistical Design of Thermal Protection System Using Physics-Informed Neural Network

T0 review · 5 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 Replacing finite-difference heat solvers with a physics-informed neural network, this paper shows that reliability-constrained thermal protection system designs can be produced in about 0.447 seconds rather than hours.

desk verdict Useful speed demo, but the statistical core doesn't identify ρ and cp separately—only k and ρ·cp are recoverable from back-temperature data. read the letter →

arxiv 2501.18078 v1 pith:4YAKN4QR submitted 2025-01-30 cs.CE

Karthik Reddy Lyathakula , Aseem Muhammad , Sevki Cesmeci This is my paper · ORCID
classification cs.CE
keywords thermalprotectionsystemphysics-informedneuralnetworksequentialMonteCarlouncertaintyquantificationreliability-baseddesignheatconductionparameterestimationsurrogatemodeling
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 aims to show that reliability-based design of a spacecraft thermal protection system can be done almost instantly: a physics-informed neural network (PINN) trained on the one-dimensional heat equation stands in for the traditional numerical solver, and a sequential Monte Carlo (SMC) sampler draws the material parameter sets that keep the back temperature below a critical limit at a chosen reliability score. The payoff, if the claim holds, is that designers can explore many materials, thicknesses, and reliability levels in seconds rather than waiting hours per run. The reported speed-up is about 175 times for 10,000 samples compared with serial MCMC on a numerical model, with the full sampling run taking 0.447 seconds. The paper also reports that PINN accuracy is comparable to the finite-difference solution, with a root-mean-square error of 3.43 degrees Celsius over the space-time domain.

What carries the argument

The two load-bearing pieces are the PINN surrogate and the SMC sampler. The PINN embeds the residual of the 1D heat equation, $f(x,t)=\frac{\partial u}{\partial t} - \frac{k}{\rho c_p}\frac{\partial^2 u}{\partial x^2}$, into its loss function, so it can predict the full temperature field for any parameter set without solving a PDE each time; its output feeds the likelihood in a Bayesian formulation. The SMC sampler uses a sequence of tempered distributions $\pi_t(\theta \mid D) \propto \pi(D\mid\theta)^{\phi_t}\pi(\theta)$, with $\phi$ going from 0 to 1, weight updates, and resampling by parallel MCMC moves when the effective sample size drops. This parallel structure is what converts the per-run cost of a numerical solver into a nearly batch-constant surrogate cost.

What would settle it

Generate synthetic back-temperature histories from two materials with the same k/(rho*cp) but very different rho and cp, run the SMC sampler, and check whether the posterior separates the two cases; if the posteriors overlap completely, the separate estimates of rho and cp are not identified and the independent-parameter treatment is not supported by the data.

Watch

Extended reading notes

Core claim

The central claim is that the combination of a PINN surrogate and SMC sampling can approximate the posterior distribution of thermal material parameters that satisfy a TPS reliability constraint, and can do so faster than MCMC coupled with numerical solvers. The PINN is trained with a loss combining physics residuals, initial condition, and boundary conditions for 1D heat conduction, and accepts k, rho, cp (with rho and cp entering as the product rho*cp) plus geometry and time as inputs. SMC then moves a population of particles through tempered distributions from prior to posterior, with parallel MCMC kernels used for resampling. The paper demonstrates target back-temperature distributions for 95%, 99%, and 99.999% reliability and shows how the implied thermal diffusivity and conductivity distributions shift with the reliability goal.

Load-bearing premise

The back temperature depends on rho and cp only through the product rho*cp (and on k through k/(rho*cp)), so the framework cannot separately identify density and heat capacity even though it assigns independent priors to each material parameter.

Editorial extensions

If this is right

  • For a fixed reliability score, the design constraint translates into a target back-temperature distribution; higher reliability (99.999% versus 95%) shifts the required material property distributions, such as lower mean back temperature and higher thermal diffusivity for a given conductivity.
  • Because PINN inference can be parallelized, running thousands of parameter samples costs about the same as running one, so the main computational bottleneck moves from the simulation itself to PINN training.
  • Material constraints, such as a cap on thermal conductivity, appear as sharp cutoffs in the feasible parameter region, allowing designers to read off diffusivity limits implied by available materials.
  • The same trained PINN can be evaluated against different reliability targets without retraining, making multi-scenario design studies inexpensive.
  • The 175x speed-up from SMC over serial MCMC is realized only when all 10,000 particles are run in parallel, which is the intended mode of use for GPU or multi-core hardware.

Reading between the lines

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

  • Because the model depends on rho and cp only through the product rho*cp, a more honest parameterization would sample theta = [k, rho*cp] or the thermal diffusivity alpha = k/(rho*cp) rather than treating rho and cp as independent; this would remove the identifiability issue without changing the reliability predictions.
  • The same PINN-plus-SMC pattern should transfer to multi-layer TPS stacks or temperature-dependent material properties, where the surrogate would need additional inputs but the sampling machinery would remain unchanged.
  • If the reported 0.447-second wall-clock time excludes PINN training, the practical advantage only fully materializes when many design evaluations share one trained surrogate; for a single one-off design, the training cost could dominate the comparison.
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

5 major / 5 minor

Summary. The paper proposes a computational framework for statistical design of thermal protection systems (TPS). The core idea is to replace a finite-difference solution of the one-dimensional heat equation with a physics-informed neural network (PINN) surrogate and then use sequential Monte Carlo (SMC) sampling to obtain distributions of material parameters that are consistent with target back-temperature reliability requirements. The authors validate the PINN against a finite-difference solver, generate target back-temperature distributions for reliability scores of 95%, 99%, and 99.999%, and report posterior-like distributions of thermal properties along with substantial computational speed-ups, including a claim that 10,000 SMC samples can be processed in 0.447 seconds versus 5–8 hours for MCMC with numerical solvers.

Significance. The practical motivation is strong: rapid reliability-based TPS design is genuinely useful, and the combination of a PINN surrogate with parallel SMC sampling is a sensible idea for reducing turnaround time. If the statistical formulation were correct, the reported speed-up would be valuable for design-space exploration. The paper also makes a concrete comparison of PINN accuracy against a finite-difference baseline, which is a useful engineering data point. However, the methodological core has serious gaps: the likelihood function is misspecified, the relationship between the reliability score and the posterior is not derived, and the material parameters ρ and cp are not identifiable from back-temperature observations alone. These issues directly undermine the paper's central claim of estimating thermal material parameter distributions.

major comments (5)
  1. [Section 4.1, Eq. (7)] The likelihood formula is not a valid Gaussian density: it is missing the negative sign in the exponent and the exponent n is undefined. As written, π(D|θ) increases without bound as the discrepancy between predicted and critical temperature grows, which is the opposite of a likelihood. This error alone invalidates the Bayesian update used in both MCMC and SMC.
  2. [Sections 4.1–4.2, Eqs. (7)–(8)] The paper does not derive how the Gaussian target distribution in Eq. (8) follows from the likelihood in Eq. (7) and the prior. The reliability score R is used to set the mean of a Gaussian target for T_back, and then MCMC/SMC samples are drawn to match that target. There is no demonstration that the resulting posterior is the correct conditional distribution given the design constraint T_back < T_crit. The target distribution is essentially an input chosen by the user, so the posterior over θ is a function of the chosen μ_target and σ_target rather than a direct consequence of the reliability constraint.
  3. [Section 3, Eq. (1) and Section 5] The model is structurally non-identifiable for ρ and cp individually. The 1D heat equation depends on material properties only through α = k/(ρcp), and the flux boundary condition involves Q/k, so the back temperature T_back(θ) depends only on k and the product ρcp. The paper nevertheless defines θ = [k, ρ, cp] with independent priors and later reports only thermal density (ρcp), thermal diffusivity, and k in Figure 8. Consequently, any posterior information about ρ and cp separately comes entirely from the priors, not from the reliability constraint. The claimed estimation of θ = [k, ρ, cp] is therefore unsupported; only k and ρcp are identifiable.
  4. [Section 5, Figure 8a] The observation that higher reliability scores lead to lower mean back temperatures is a restatement of the construction in Eq. (8), where R directly determines the mean of the target back-temperature distribution. This is not an independent finding of the framework, and it cannot be used to argue that the method 'satisfies' the reliability constraints; by construction the samples are drawn to match the target distribution.
  5. [Section 5, computational comparison] The reported speed-up is not a like-for-like comparison. The 0.447-second runtime for 10,000 SMC samples excludes PINN training time, while the 5–8 hour estimate for MCMC with numerical solvers appears to include full forward solves. The 175× speed-up also conflates two changes: replacing the numerical solver with a PINN and parallelizing SMC versus running MCMC serially. A fair benchmark should state hardware, include all preprocessing/training cost, and isolate the contribution of each component.
minor comments (5)
  1. [Section 3, Eq. (5)] Both terms in the boundary condition loss are written with u_θ(0, t_i), but the second term should correspond to x = L where the heat flux is applied; as written the formula is inconsistent with the stated boundary conditions.
  2. [Section 5] Several equations and tables are referenced as 'Eq x' or 'table x' without numbers, and several figures lack complete captions; this makes it difficult to reproduce or check the results.
  3. [Section 4.3] The weight update formula in the SMC description is missing a summation in the denominator index notation and could be clarified; the current formula appears to have an indexing inconsistency between w_i and w_j.
  4. [Section 4.2] The statement that all burn-in samples are included in the uncertainty quantification is unusual and may bias the reported distributions; the authors should justify this choice or exclude burn-in as is standard.
  5. [References] There are minor typographical issues in the references, such as 'Howel' for 'Howell', and some references are incomplete (e.g., the SMCPy entry).

Circularity Check

2 steps flagged · score 4.0 of 10

Core speed/accuracy claims are externally benchmarked, but the reliability-trend result is built into Eq. 8 and the separate rho/cp posteriors reduce to priors.

  1. fitted input called prediction [Section 4.2 (Eq. 8) and Section 5, Figure 8a discussion]
    "With this assumption, reliability score R is used to find the z-score of the gaussian distribution, which in turn, is used to find the mean of the distribution. The reliability score is defined as: R=P(T_back ≤ T_critical)=N((T_critical−μ_target)/σ_target) ... It can be observed that the target distributions have lower mean back temperature as the reliability score increases, as expected."

    The mean μ_target is solved from R and σ_target in Eq. 8, so for fixed σ_target, increasing R mechanically shifts μ_target downward. Figure 8a's 'observation' is therefore the defining formula restated, not a result of PINN/SMC estimation. The reliability requirement is an input; the back-temperature distribution is calibrated to it. This is a fitted-input-called-prediction: the target distribution is constructed from R and then reported as if the framework discovered that higher reliability implies lower mean back temperature.

  2. self definitional [Eq. 1 (governing PDE), Section 4.1 (parameter set), Section 5 (PINN inputs)]
    "∂T/∂t = k/(ρ c_p) ∂²T/∂x² ... The inputs to PINN are the points in the physics domain, points on the boundary and initial conditions, and the parameters [k, ρ, c_p]. In this implementation, ρ and c_p are combined to find the thermal density (ρ∗c_p) of the TPS which is then used as the input."

    The PDE (Eq. 1) and the boundary condition ∂T/∂x|_L=−Q/k contain ρ and c_p only as the product ρc_p. T_back(θ), and therefore the likelihood π(D|θ) in Eq. 7, are invariant along the ridge ρ·c_p=const. Hence the posterior over θ=[k,ρ,c_p] identifies only k and ρc_p; any separate marginal for ρ or c_p is the input prior π(ρ)π(c_p), unchanged by the reliability constraint. The claimed estimation of θ's individual components thus reduces, for ρ and c_p, to the prior by construction.

full rationale

Score 4: The PINN-vs-FDM accuracy check, the wall-clock speed comparisons, and the SMC-vs-MCMC speedup are external benchmarks with independent content, so the core computational claims are not circular. Self-citations [22], [28], [37] are background and not load-bearing. Two partial circularities remain. First, the reliability trend 'higher R ⇒ lower mean back temperature' is imposed by Eq. 8, which defines the target mean from R and σ_target; Figure 8a restates that construction. Second, because the governing equation and boundary condition depend on material properties only through k and ρc_p, the posterior for ρ and c_p individually is prior-dominated; the paper nevertheless defines θ=[k,ρ,c_p] and describes estimating thermal material parameters, although its own plots report only identifiable combinations. These issues are not central to the speed and surrogate-accuracy findings, but they do weaken the 'parameter estimation' framing.

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

The central claim rests on the choice of a Gaussian target for back temperature, a Gaussian likelihood, and the non-identifiable treatment of rho and cp. The surrogate validation and speed benchmark provide external grounding, but the statistical design results depend on several unquantified user choices.

free parameters (5)
  • sigma_target = not specified
    Standard deviation of the assumed Gaussian target back-temperature distribution in Eq. 8; chosen by the user and controls the width of the resulting parameter distributions.
  • prior means and standard deviations for rho, cp, k = not specified
    Initial normal distributions used to seed MCMC/SMC; not reported, but they influence the posterior through Eq. 7.
  • thermal conductivity upper bound = 1 W/mK
    A self-imposed limit mentioned in the results as creating the sharp line in Figure 8d; not justified from material data.
  • PINN loss weights alpha_1, alpha_2, alpha_3 = not specified
    Weights in Eq. 6 balancing physics, initial, and boundary losses; values not reported.
  • PINN training parameter range = wide range centered on RCC material
    The surrogate is only valid inside this range; boundaries are not quantified.
assumptions (5)
  • domain assumption One-dimensional heat conduction with constant thermal properties (Eq. 1) is an adequate model for TPS back temperature during re-entry.
    Radiation and lateral heat conduction are explicitly ignored (Section 3).
  • domain assumption Adiabatic back surface and constant heat flux at the outer surface with Q=10000 W/m^2 and L=7mm are representative boundary conditions.
    Section 5 sets these values; no sensitivity to these choices is reported.
  • ad hoc to paper The discrepancy between predicted back temperature and critical temperature is Gaussian with standard deviation sigma (Eq. 7).
    This is a statistical assumption introduced without empirical support.
  • ad hoc to paper The target back-temperature distribution is Gaussian and its mean is determined by the reliability score R and sigma_target (Eq. 8).
    The Gaussian form and sigma_target are chosen, not derived from data or physics.
  • domain assumption The priors on rho, cp, and k are independent.
    Section 4.1 writes pi(theta)=pi(rho)pi(cp)pi(k), despite the PDE depending only on the product rho*cp.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Statistical Design of Thermal Protection System Using Physics-Informed Neural Network." pith.science (2026). https://pith.science/paper/4YAKN4QR

@misc{pith2026250118078,
  author       = {Pith},
  title        = {Pith review of: Statistical Design of Thermal Protection System Using Physics-Informed Neural Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4YAKN4QR}},
  note         = {Machine review of arXiv:2501.18078}
}
read the original abstract

Thermal protection systems (TPS) of space vehicles are designed computationally rather than experimentally. They are validated using ground experiments, but all aspects of the flight cannot be replicated on ground. This ground-to-flight mapping introduces uncertainties which need to be accounted for while designing any thermal protection system. Thus, precise computational models along with uncertainty quantification in the models are required to design the TPS. The focus of this study is to estimate the thermal material parameters of TPS based on the target reliability requirements using statistical methods. To perform uncertainty quantification (UQ) of a system, a simulated model of the system needs to be solved many times on statistical samples, increasing the computational time and cost of the overall process. A physics-informed neural network (PINN) model is used in the analysis instead of traditional physics based numerical solutions. The accuracy of PINN is comparable to that of the numerical solution. To find the parameter distribution, sampling of the parameter space is performed using Sequential Monte- Carlo (SMC) method. The sampling method is efficient as it generates samples based on the target distribution in parallel and it also generates diverse samples for proper UQ. Combining the use of both PINN predictive model and SMC sampling, the framework can approximate the parameter distributions that satisfy the TPS design reliability constraints. The framework achieved remarkable increases in the speed of performing the reliability analysis of the TPS. This reliability analysis can be used for design optimization of the TPS based on risk analysis along with other systems of the vehicle.

Figures

Figures reproduced from arXiv: 2501.18078 by the authors.

Figure 1
Figure 1. Probabilistic Design Process 3. Predictive Physics Model The TPS of space vehicles experiences extreme heat flux during re-entry because of the air friction in the lower atmosphere. The back temperature should not exceed a certain limit on the inside of the TPS. For this study, heat transfer by radiation as well as heat transfer in longitudinal directions will be ignored. Thus, the simplified thermal response of a T… view at source ↗
Figure 2
Figure 2. a) Thermal Protectional System of Space X's Starship [29] b) Thermal Model for TPS [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Schematic of a physics-informed neural network (PINN) [5] [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Markov Chain Monte-Carlo Process Diagram[37] [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Sequential Monte Carlo Process [37] 5. Results and Discussion This section presents the results of the methodology presented in section 2. The numerical solution of the Eq. 2 is implemented in Python programming language using finite difference method (FDM) technique. …
Figure 6
Figure 6. Figure 6: Loss Curve of PINN for the validation case [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: a) PINN Prediction b) Numerical Solution c) Absolute Error between PINN and Numerical Solution [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: a) Target Back Temperature Distributions with Different Reliability Scores b) Probability Density of Thermal Density [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: a) PINN Inference Time vs Numerical Solution Time. Constant [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 48 canonical work pages

  1. [5]

    Create a PINN Surrogate Model: Design a neural network that incorporates the rules and equations from the numerical model to guide its learning process

    work page

  2. [37]

    PINN surrogate of Li-ion battery models for parameter inference. Part I: Implementation and multi-fidelity hierarchies for the single-particle model

    M. Hassanaly et al., “PINN surrogate of Li-ion battery models for parameter inference. Part I: Implementation and multi-fidelity hierarchies for the single-particle model,” Sep. 09, 2024, arXiv: arXiv:2312.17329. doi: 10.48550/arXiv.2312.17329

    work page Pith review arXiv doi:10.48550/arxiv.2312.17329 2024

  3. [1]

    critical temperatures

    Introduction During the re-entry stage of a space vehicle, a significant amount of heat is generated due to air friction. This heat increases the temperature of the surface of the space vehicle, more in some areas than others. A heat shield or thermal protection system is designed to prevent any damage to the internal structure of the space vehicle. Diffe...

    work page

  4. [2]

    Methodology The methodology of the probabilistic design optimization process is as follows:

    work page

  5. [3]

    Train and Test the PINN Model: Train the numerical model to make accurate predictions and compare its results with test data to ensure it works correctly

    work page

  6. [4]

    critical temperature for the bottom side of the TPS

    Set Up a Numerical Model: Build a mathematical model that represents how the thermal protection system (TPS) works and apply the necessary conditions at the edges i.e. critical temperature for the bottom side of the TPS

    work page

  7. [6]

    Analyze Uncertainty and Reliability: Study the range of possible values for the key parameters and calculate how reliable the design is based on the results

    work page

  8. [7]

    Define Limits and Reliability Goals: Set the maximum allowable temperature critical temperature on the bottom of the TPS, decide on a reliability score to measure success, and define initial guesses for the key parameters to be optimized

    work page

Show all 55 references
  1. [8]

    Keep sampling until the parameters stabilize and meet the critical temperature constraint

    Use MCMC for Parameter Sampling: Apply Markov Chain Monte Carlo (MCMC) to explore different combinations of the key parameters. Keep sampling until the parameters stabilize and meet the critical temperature constraint

  2. [9]

    Needs and Opportunities for Uncertainty-Based Multidisciplinary Design Methods for Aerospace Vehicles,

    T. A. Zang et al., “Needs and Opportunities for Uncertainty-Based Multidisciplinary Design Methods for Aerospace Vehicles,” NAS 1.15:211462, Jul. 2002. Accessed: Jan. 04, 2025. [Online]. Available: https://ntrs.nasa.gov/citations/20020063596

    2002

  3. [10]

    Adjust the parameters and constraints to improve performance if needed

    Perform Sensitivity Analysis: Test how changes in different parameters affect the model. Adjust the parameters and constraints to improve performance if needed

  4. [11]

    If not, continue adjusting and testing

    Finalize the Design: If the parameters meet the design goals and stay within the constraints, stop the process. If not, continue adjusting and testing. Figure 1: Probabilistic Design Process

  5. [12]

    The back temperature should not exceed a certain limit on the inside of the TPS

    Predictive Physics Model The TPS of space vehicles experiences extreme heat flux during re-entry because of the air friction in the lower atmosphere. The back temperature should not exceed a certain limit on the inside of the TPS. For this study, heat transfer by radiation as ...

  6. [13]

    The process begins with the selection of initial parameters, where geometric dimensions and material properties are chosen based on available materials or engineering estimates

    Parameter Estimation and Uncertainty Quantification Uncertainty quantification (UQ) in engineering systems typically relies on the Monte Carlo method, which follows a structured three-step approach [1]. The process begins with the selection of initial parameters, where geometr...

  7. [14]

    The numerical solution of the Eq

    Results and Discussion This section presents the results of the methodology presented in section 2. The numerical solution of the Eq. 2 is implemented in Python programming language using finite difference method (FDM) technique. In this study, the heat flux is set to 10,000 𝑊...

    2000

  8. [15]

    The PINN model showed good accuracy while running many times faster for parallel simulations

    Conclusion This study investigated a new approach to designing thermal protection systems by combining physics- informed neural networks with sequential Monte Carlo sampling. The PINN model showed good accuracy while running many times faster for parallel simulations. Using SM...

  9. [16]

    Probabilistic Design Method for Aircraft Thermal Protective Layers Based on Surrogate Models,

    Z. Chen, K. Zhang, S. Zhao, F. Li, F. Xu, and M. Chen, “Probabilistic Design Method for Aircraft Thermal Protective Layers Based on Surrogate Models,” Energies, vol. 17, no. 5, Art. no. 5, Jan. 2024, doi: 10.3390/en17051051

    2024 doi

  10. [17]

    Efficient strategy for reliability-based optimization design of multidisciplinary coupled system with interval parameters,

    R. Wang and Y . Luo, “Efficient strategy for reliability-based optimization design of multidisciplinary coupled system with interval parameters,” Appl. Math. Model., vol. 75, pp. 349–370, Nov. 2019, doi: 10.1016/j.apm.2019.05.030

    2019 doi

  11. [18]

    Design of Thermal Protection System for Reusable Hypersonic Vehicle Using Inverse Approach,

    S. Kumar and S. P. Mahulikar, “Design of Thermal Protection System for Reusable Hypersonic Vehicle Using Inverse Approach,” J. Spacecr. Rockets, vol. 54, no. 2, pp. 436–446, 2017, doi: 10.2514/1.A33688

    2017 doi

  12. [19]

    Thermal Protection System and Thermal Management for Combined-Cycle Engine: Review and Prospects,

    Y . Dong, E. Wang, Y . You, C. Yin, and Z. Wu, “Thermal Protection System and Thermal Management for Combined-Cycle Engine: Review and Prospects,” Energies, vol. 12, no. 2, Art. no. 2, Jan. 2019, doi: 10.3390/en12020240

    2019 doi

  13. [20]

    A Parametric Physics-Informed Deep Learning Method for Probabilistic Design of Thermal Protection Systems,

    R. Zhang, N. Xu, K. Zhang, L. Wang, and G. Lu, “A Parametric Physics-Informed Deep Learning Method for Probabilistic Design of Thermal Protection Systems,” Energies, vol. 16, no. 9, Art. no. 9, Jan. 2023, doi: 10.3390/en16093820

    2023 doi

  14. [21]

    Thermal protection systems for space vehicles: A review on technology development, current challenges and future prospects,

    O. Uyanna and H. Najafi, “Thermal protection systems for space vehicles: A review on technology development, current challenges and future prospects,” Acta Astronaut., vol. 176, pp. 341–356, Nov. 2020, doi: 10.1016/j.actaastro.2020.06.047

    2020 doi

  15. [22]

    Probabilistic transient thermal analysis of an atmospheric reentry vehicle structure,

    T. Nakamura and K. Fujii, “Probabilistic transient thermal analysis of an atmospheric reentry vehicle structure,” Aerosp. Sci. Technol., vol. 10, no. 4, pp. 346–354, May 2006, doi: 10.1016/j.ast.2006.02.002

    2006 doi

  16. [23]

    C. J. Katsikas, G. K. Castle, and J. S. Higgins, Ablation Handbook Entry Materials Data and Design. Air Force Materials Laboratory, Research and Technology Division, Air Force Systems Command, 1966

    1966

  17. [24]

    Monte Carlo Treatment of Data Uncertainties in Thermal Analysis,

    J. R. Howell, “Monte Carlo Treatment of Data Uncertainties in Thermal Analysis,” J. Spacecr. Rockets, vol. 10, no. 6, pp. 411–414, Jun. 1973, doi: 10.2514/3.61899

    1973 doi

  18. [25]

    and Ravishankar [15] also presented uncertainty analysis of aerothermal and thermal protection material responses using the Monte Carlo method. This technique provides a risk-based probabilistic analysis approach, whereby the TPS can be designed at the desired risk level, and ...

  19. [26]

    Reliability-Based Optimization for Multidisciplinary System Design,

    D. Padmanabhan, “Reliability-Based Optimization for Multidisciplinary System Design,” thesis, University of Notre Dame, 2003. doi: 10.7274/2r36tx33k39

    2003 doi

  20. [27]

    Probabilistic Design of a Mars Sample Return Earth Entry Vehicle Thermal Protection System,

    J. A. Dec and R. A. Mitcheltree, “Probabilistic Design of a Mars Sample Return Earth Entry Vehicle Thermal Protection System,” Jan. 01, 2002. Accessed: Jan. 04, 2025. [Online]. Available: https://ntrs.nasa.gov/citations/20020022116

    2002

  21. [28]

    R. C. Smith, Uncertainty Quantification: Theory, Implementation, and Applications, Second Edition. Philadelphia, PA: Society for Industrial and Applied Mathematics, 2024. doi: 10.1137/1.9781611977844

    2024 doi

  22. [29]

    Heat Conduction Plate Layout Optimization Using Physics-Driven Convolutional Neural Networks,

    Y . Sun, A. Elhanashi, H. Ma, and M. R. Chiarelli, “Heat Conduction Plate Layout Optimization Using Physics-Driven Convolutional Neural Networks,” Appl. Sci., vol. 12, no. 21, Art. no. 21, Jan. 2022, doi: 10.3390/app122110986

    2022 doi

  23. [30]

    On Machine-Learning-Driven Surrogates for Sound Transmission Loss Simulations,

    B. Z. Cunha, A.-M. Zine, M. Ichchou, C. Droz, and S. Foulard, “On Machine-Learning-Driven Surrogates for Sound Transmission Loss Simulations,” Appl. Sci., vol. 12, no. 21, Art. no. 21, Jan. 2022, doi: 10.3390/app122110727

    2022 doi

  24. [31]

    Uncertainty Analysis of Integrated Thermal Protection System with Rigid Insulation Bars,

    B. Ravishankar, R. Haftka, and B. Sankar, “Uncertainty Analysis of Integrated Thermal Protection System with Rigid Insulation Bars,” in 52nd AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference, Denver, Colorado: American Institute of Aeronautics and...

    2011 doi

  25. [32]

    Uncertainty Quantification of Aero-Thermal Performance of a Blade Endwall Considering Slot Geometry Deviation and Mainstream Fluctuation,

    Z. Tao, Z. Guo, L. Song, and J. Li, “Uncertainty Quantification of Aero-Thermal Performance of a Blade Endwall Considering Slot Geometry Deviation and Mainstream Fluctuation,” J. Turbomach., vol. 143, no. 111013, Jun. 2021, doi: 10.1115/1.4051416

    2021 doi

  26. [33]

    Thermo-mechanical optimization of metallic thermal protection system under aerodynamic heating,

    Q. Guo, S. Wang, W. Hui, Y . Li, and Z. Xie, “Thermo-mechanical optimization of metallic thermal protection system under aerodynamic heating,” Struct. Multidiscip. Optim., vol. 61, no. 2, pp. 819– 836, Feb. 2020, doi: 10.1007/s00158-019-02379-4

    2020 doi

  27. [34]

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

    M. Raissi, P . Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,” J. Comput. Phys., vol. 378, pp. 686–707, Feb. 2019, doi: 10.1016/j.jcp.2...

    2019 doi

  28. [35]

    Towards a Hybrid Digital Twin: Physics-Informed Neural Networks as Surrogate Model of a Reinforced Concrete Beam,

    T. Sahin, D. Wolff, M. von Danwitz, and A. Popp, “Towards a Hybrid Digital Twin: Physics-Informed Neural Networks as Surrogate Model of a Reinforced Concrete Beam,” in 2024 Sensor Data Fusion: Trends, Solutions, Applications (SDF), Nov. 2024, pp. 1–8. doi: 10.1109/SDF63218.202...

    2024

  29. [36]

    A PINN Surrogate Modeling Methodology for Steady- State Integrated Thermofluid Systems Modeling,

    K. Laugksch, P. Rousseau, and R. Laubscher, “A PINN Surrogate Modeling Methodology for Steady- State Integrated Thermofluid Systems Modeling,” Math. Comput. Appl., vol. 28, no. 2, Art. no. 2, Apr. 2023, doi: 10.3390/mca28020052

    2023 doi

  30. [38]

    Physics-Informed Deep Learning- Based Modeling of a Novel Elastohydrodynamic Seal for Supercritical CO2 Turbomachinery,

    K. R. Lyathakula, S. Cesmeci, M. DeMond, H. Xu, and J. Tang, “Physics-Informed Deep Learning- Based Modeling of a Novel Elastohydrodynamic Seal for Supercritical CO2 Turbomachinery,” in ASME Power Conference, American Society of Mechanical Engineers, 2022, p. V001T15A008. Acce...

    2022

  31. [39]

    Apollo thermal- protection system development,

    R. B. Erb, D. H. Greenshields, L. T. Chauvin, J. E. Pavlosky, and C. L. Statham, “Apollo thermal- protection system development,” J. Spacecr. Rockets, vol. 7, no. 6, pp. 727–734, Jun. 1970, doi: 10.2514/3.30027

    1970 doi

  32. [40]

    Probabilistic Modeling of Aerothermal and Thermal Protection Material Response Uncertainties,

    M. J. Wright, D. Bose, and Y .-K. Chen, “Probabilistic Modeling of Aerothermal and Thermal Protection Material Response Uncertainties,” AIAA J., vol. 45, no. 2, pp. 399–410, Feb. 2007, doi: 10.2514/1.26018

    2007 doi

  33. [41]

    A survey of Monte Carlo methods for parameter estimation,

    D. Luengo, L. Martino, M. Bugallo, V . Elvira, and S. Särkkä, “A survey of Monte Carlo methods for parameter estimation,” EURASIP J. Adv. Signal Process., vol. 2020, no. 1, p. 25, May 2020, doi: 10.1186/s13634-020-00675-6

    2020 doi

  34. [42]

    Efficient Sequential Monte-Carlo Samplers for Bayesian Inference,

    T. L. T. Nguyen, F. Septier, G. W. Peters, and Y . Delignon, “Efficient Sequential Monte-Carlo Samplers for Bayesian Inference,” Apr. 22, 2015, arXiv: arXiv:1504.05753. doi: 10.48550/arXiv.1504.05753

    2015 doi arXiv

  35. [43]

    A probabilistic fatigue life prediction for adhesively bonded joints via ANNs-based hybrid model,

    K. R. Lyathakula and F.-G. Yuan, “A probabilistic fatigue life prediction for adhesively bonded joints via ANNs-based hybrid model,” Int. J. Fatigue, vol. 151, p. 106352, Oct. 2021, doi: 10.1016/j.ijfatigue.2021.106352

    2021

  36. [44]

    SpaceX to use LG Energy batteries for Starship rocket,

    K. E. D. Global, “SpaceX to use LG Energy batteries for Starship rocket,” KED Global. Accessed: Jan. 18, 2025. [Online]. Available: https://www.kedglobal.com/batteries/newsView/ked202411110004

    2025

  37. [45]

    J. D. Hoffman, Numerical methods for engineers and scientists, 2nd ed., rev.Expanded. New York: Marcel Dekker, 2001

    2001

  38. [46]

    Implicit-explicit Runge-Kutta methods for time- dependent partial differential equations,

    U. M. Ascher, S. J. Ruuth, and R. J. Spiteri, “Implicit-explicit Runge-Kutta methods for time- dependent partial differential equations,” Appl. Numer. Math., vol. 25, no. 2, pp. 151–167, Nov. 1997, doi: 10.1016/S0168-9274(97)00056-1

    1997 doi

  39. [47]

    Automatic differentiation in machine learning: a survey,

    A. G. Baydin, B. A. Pearlmutter, A. A. Radul, and J. M. Siskind, “Automatic differentiation in machine learning: a survey,” Feb. 05, 2018, arXiv: arXiv:1502.05767. doi: 10.48550/arXiv.1502.05767

    2018 doi arXiv

  40. [48]

    Deep learning | Nature

    “Deep learning | Nature.” Accessed: Jan. 27, 2025. [Online]. Available: https://www.nature.com/articles/nature14539

    2025

  41. [49]

    A physics-informed deep learning framework for inversion and surrogate modeling in solid mechanics - ScienceDirect

    “A physics-informed deep learning framework for inversion and surrogate modeling in solid mechanics - ScienceDirect.” Accessed: Jan. 27, 2025. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0045782521000773?casa_token=YaUdwpvh2iQ AAAAA:xyRLE9fLSlOsW_ES...

    2025

  42. [50]

    Deep Sparse Rectifier Neural Networks,

    X. Glorot, A. Bordes, and Y . Bengio, “Deep Sparse Rectifier Neural Networks,” in Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, JMLR Workshop and Conference Proceedings, Jun. 2011, pp. 315–323. Accessed: Jan. 07, 2025. [Onlin...

    2011

  43. [51]

    The Metropolis-Hastings algorithm,

    C. P. Robert, “The Metropolis-Hastings algorithm,” Jan. 27, 2016, arXiv: arXiv:1504.01896. doi: 10.48550/arXiv.1504.01896

    2016 doi arXiv

  44. [52]

    A Cloud-based Real-time Probabilistic Remaining Useful Life (RUL) Estimation using the Sequential Monte Carlo (SMC) Method,

    K. R. Lyathakula and F.-G. Yuan, “A Cloud-based Real-time Probabilistic Remaining Useful Life (RUL) Estimation using the Sequential Monte Carlo (SMC) Method,” Nov. 28, 2024, arXiv: arXiv:2411.17824. doi: 10.48550/arXiv.2411.17824

    2024 doi arXiv

  45. [53]

    Sequential monte carlo: Enabling real-time and high-fidelity prognostics,

    P. E. Leser, J. D. Hochhalter, J. E. Warner, G. F. Bomarito, W. P. Leser, and F.-G. Yuan, “Sequential monte carlo: Enabling real-time and high-fidelity prognostics,” in Annual Conference of the PHM Society, PHM Soc. Philadelphia, PA, 2018. Accessed: Jan. 27, 2025. [Online]. Av...

    2018

  46. [54]

    Thermal Characterization of Carbon Fiber-Reinforced Carbon Composites (C/C),

    J. D. Macías et al., “Thermal Characterization of Carbon Fiber-Reinforced Carbon Composites (C/C),” Appl. Compos. Mater., vol. 26, no. 1, pp. 321–337, Feb. 2019, doi: 10.1007/s10443-018-9694- 0

    2019 doi

  47. [55]

    Patrick, Leser and Michael, Wang, SMCPy - Sequential Monte Carlo with Python

Pith tools

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