Bayesian parameter inference for PICA devolatilization pyrolysis at high heating rates
Joffrey Coheur
Promotors:M. Arnst1, P. Chatelain2, T. Magin4
Collaborators:P. Schrooyen3, K. Hillewaert3, A. Turchi4,
F. Panerai5, J. Meurisse5, N. N. Mansour.5 1Aerospace and Mechanical Engineering, Universit´e de Li`ege
2Institute of Mechanics, Materials and Civil Engineering, Universit´e Catholique de Louvain 3Cenaero, Gosselies
4Aeronautics and Aerospace Department, von Karman Institute for Fluid Dynamics 5NASA Ames Research Center
3rd International Conference on Uncertainty Quantification in Computational Sciences and Engineering
Motivation: atmospheric entry
Thermal protection systems (TPS)
Dragon capsule (Space X)
Space debris
Ablative materials for thermal protection systems
I PICA: Phenolic-Impregnated Carbon Ablator
Mars Science Laboratory thermal protection system (NASA)
Porous thermal protection material [Helber, 2016] Bottom view (after burn) Lawson et al. Fibers + resin 2 / 22
Modeling the pyrolysis of ablative thermal protection materials
Aerodynamic simulation (CooLFluiD, VKI) Aerothermal simulation (Argo, Cenaero) [Coheur, Schrooyen, 2017] High T°gas flow Pyrolysis
gas Ablation
I CFD codes require accurate models for ablation [Lachaud, 2014; Schrooyen, 2016], e.g. conservation of mass species
∂ghρiig
∂t +∇ · (ghρiighuig) =∇ · hJii +h ˙ω pyro
i i
I Objectives: deduce h ˙ωipyroi from dedicated pyrolysis experiments
Table of Contents
Introduction
Characterization of physico-chemical parameters relevant to pyrolysis reactions Experiments
Modeling Challenges
Non-linear posterior and highly-correlated parameters Zero-variance data in the dataset
Application to complex pyrolysis model Conclusions
Table of Contents
Introduction
Characterization of physico-chemical parameters relevant to pyrolysis reactions Experiments
Modeling Challenges
Non-linear posterior and highly-correlated parameters Zero-variance data in the dataset
Application to complex pyrolysis model Conclusions
Pyrolysis experiments [Wong et al., 2015; Bessire and Minton, 2017]
Phenolic sample (50 mg) Pyrolysis gases Heated wall 400 600 800 1000 1200 0 0.2 0.4 0.6 0.8 Temperature, T Mass yields, mg H2 CO CO2 CH4 H2O 400 600 800 1000 1200 30 40 50 Temperature, K Sample mass, mg 4 / 22Table of Contents
Introduction
Characterization of physico-chemical parameters relevant to pyrolysis reactions Experiments
Modeling Challenges
Non-linear posterior and highly-correlated parameters Zero-variance data in the dataset
Application to complex pyrolysis model Conclusions
Phenomenological laws for pyrolysis mass loss and gas species production
I Pyrolysis decomposition of ablative materials follows successive reaction rates [Goldstein, 1969; Trick, 1997] 400 600 800 1000 1200 0 2 4 6 8 ·10−3 Temperature, K ∂ ξj /∂ T j = 1 j = 2 400 600 800 1000 1200 0.4 0.6 0.8 1 Temperature, K m / m0 h ˙ωpyroi i= Np X j Fij∂ξj ∂Tτ m0 ∂ξj ∂T = (1− ξj) nj Aj τ exp − Ej RT
I ξj: advancement of reaction of the fictitious resin component j
Phenomenological laws for pyrolysis mass loss and gas species production
I Pyrolysis decomposition of ablative materials follows successive reaction rates [Goldstein, 1969; Trick, 1997] 400 600 800 1000 1200 0 2 4 6 8 ·10−3 Temperature, K ∂ ξj /∂ T j = 1 j = 2 400 600 800 1000 1200 0.4 0.6 0.8 1 Temperature, K m / m0 m = m0− Np X j Fjξjm0 ∂ξj ∂T = (1− ξj) nj Aj τ exp − Ej RT
Bayesian inference for parameter calibration
Objective: Infer on a finite set of parameters p∈ Rq from
I A set of measurements dobs={d
i ∈ R, i = 1, . . . , nobs}
I A model that predicts the measurements
using Bayesian rule for improving our knowledge on p
π(p|dobs) = π(d obs|p)π 0(p) R Rpπ(d obs|p)π 0(p)dp (Bayes’ theorem)
I Choice for the prior π0(p): uniform pdf (bounded support)
I Choice for the likelihood π(dobs|p): additive Gaussian noise π(dobs|p) = 1 Πi(2πσi2)n/2 exp − ns X i=1 n X k=1 dobs ik − ηi(xk, p) 2 2σi2 ! 6 / 22
Benchmark: two-equation model
200 400 600 800 1,000 1,200 1,400 0 0.5 1 1.5 ·10 −3 Temperature, K Pro duction rate, s − 1 Reaction I Reaction IISynthetic data using: σI= 0.149 s−1
σII = 0.5382 s−1
A E m F
Reaction I 541218 74046.7 9.08 0.11 Reaction II 5360000 103680 5.07 0.35
Random-Walk Metropolis-Hastings: bivariate posterior PDFs
0.4 0.6 0.8 1 1.2 1.4 ·105 0 0.5 1 1.5 2 ·10 8 E A 4 6 8 10 12 14 0 0.5 1 1.5 2 ·10 8 m A 0 0.1 0.2 0.3 0.4 0 0.5 1 1.5 2 ·10 8 F A 4 6 8 10 12 14 0.4 0.6 0.8 1 1.2 1.4 ·10 5 m E 0 0.1 0.2 0.3 0.4 0.4 0.6 0.8 1 1.2 1.4 ·10 5 F E 0 0.1 0.2 0.3 0.4 4 6 8 10 12 14 F m 8 / 22Challenges for the inference of pyrolysis kinetic parameters
I Non-linear posterior and highly-correlated parameters: random-walk Metropolis-Hastings may fail or be very slow, tuning of the proposal covariance hard for high dimensional problems
• Reparametrization of parameters space adapted to the physical model
• Gradient-based algorithms (Itˆo-SDE)
I Fitting unimportant data may biased the results
• Hyperparameter choice
Table of Contents
Introduction
Characterization of physico-chemical parameters relevant to pyrolysis reactions Experiments
Modeling Challenges
Non-linear posterior and highly-correlated parameters Zero-variance data in the dataset
Application to complex pyrolysis model Conclusions
Reparametrization of parameters space adapted to the physical model
I More complex models become prohibitive in terms of number of iterations required I Need to improve the mixing of the Markov chains
Non-linear transformation of the parameter space ˜ Ai = ln Ai − Ei/(RTi), ˜ Ei = Ei/Ei, ˜ mi = mi/ ¯mi, ˜ Fi = Fij/Fij. Reaction rate: ki = exp ˜Ai + ˜EiT˜i(t)
Local reciprocal temperature: ˜
Benchmark: two-equation model
200 400 600 800 1,000 1,200 1,400 0 0.5 1 1.5 ·10 −3 Temperature, K Pro duction rate, s − 1 Reaction I Reaction IISynthetic data using: σI= 0.149 s−1 σII = 0.5382 s−1 A E m F Reaction I 541218 74046.7 9.08 0.11 Reaction II 5360000 103680 5.07 0.35 11 / 22
Bivariate posterior PDFs (initial parameter space)
0.4 0.6 0.8 1 1.2 1.4 ·105 0 0.5 1 1.5 2 ·10 8 E A 4 6 8 10 12 14 0 0.5 1 1.5 2 ·10 8 m A 0 0.1 0.2 0.3 0.4 0 0.5 1 1.5 2 ·10 8 F A 4 6 8 10 12 14 0.4 0.6 0.8 1 1.2 1.4 ·10 5 m E 0 0.1 0.2 0.3 0.4 0.4 0.6 0.8 1 1.2 1.4 ·10 5 F E 0 0.1 0.2 0.3 0.4 4 6 8 10 12 14 F mBivariate posterior PDFs (rescaled parameter space)
0.8 1 1.2 −2.6 −2.4 −2.2−2 −1.8 −1.6 ˜ E ˜ A 0.6 0.8 1 1.2 1.4 1.6 −2.6 −2.4 −2.2−2 −1.8 −1.6 ˜ m ˜ A 0.9 1 1.1 1.2 −2.6 −2.4 −2.2−2 −1.8 −1.6 ˜ F ˜ A 0.6 0.8 1 1.2 1.4 1.6 0.8 1 1.2 ˜ m ˜ E 0.9 1 1.1 1.2 0.8 1 1.2 ˜ F ˜ E 0.9 1 1.1 1.2 0.6 0.8 1 1.2 1.4 1.6 ˜ F ˜m 13 / 22Itˆ
o Stochastic Differential Equation based MCMC method
We use the Itˆo Stochastic Differential Equation (ISDE) introduced by Soize (2008) dΞ = [ bC ]Hdt dH = −∇ξφ(Ξ)dt− 1 2ζ0Hdt + pζ0LCˆ −T dW with φ =− log(π(dobs|p)), [ bC ]: approximation of the covariance matrix, ζ
0: free parameter
(Arnst, Soize (2019)). Discretization in time using a Stormer-Verlet method (Soize, Ghanem (2016)) Ξ(`+12) = Ξ(`)+4t 2 [ bC ]H(`) H(`+1)= 1−b 1+bH(`)−1+b4t ∇ξφ Ξ(`+12) + √ζ0 1+bLCb −T ∆W(`+1) Ξ(`+1)= Ξ(`+12) + 4t 2 [ bC ]H(`+1)
Table of Contents
Introduction
Characterization of physico-chemical parameters relevant to pyrolysis reactions Experiments
Modeling Challenges
Non-linear posterior and highly-correlated parameters Zero-variance data in the dataset
Application to complex pyrolysis model Conclusions
Zero variance data in the data-set
I Hyperparameter choice π(dobs|p) = 1 Πi(2πσi2)n/2 exp − ns X i=1 n X k=1 dobs ik − ηi(xk, p) 2 2σi2 !I Taking all the data might lead to inaccurate results, although zero variance data (before activation temperature of after reaction) are obtained experimentally
I Feedback on the experiments: fit only important features (activation temperature, maximum production peaks)
Illustration using experimental data: two-reaction model with one species
I Observable (data) dHobs2k: mass yields of species H2 at each temperature. Np= 2.
dH2k = Np X j FH2,jm0 ξj(k)− ξj(k−1) , ∂ξj(k) ∂t = 1− ξ(k)j nj Ajexp − Ej RTk , ξj(0)= 0. 400 600 800 1000 1200 1400 0 2 4 ·10−6 Temperature, K Pro duction rate, s − 1
I Simple point-mass model (0D), k = 1, . . . , nT
I p={A1, E1, n1, FH2,1, A2, E2, n2, FH2,2}
I Initial guess obtained from genetic algorithm
Comparison of the different approaches
I Comparison of posterior samples obtained from the Markov chains (length = 1e4) using Itˆo-SDE (blue dots) and random-walk Metropolis-Hastings in the reparametrized parameter space (orange dots) and in the initial parameter space (green dots)
0.9 1 1.1 1.2 1.3 1.4 1.5 ·105 0 0.5 1 ·107 E A
Table of Contents
Introduction
Characterization of physico-chemical parameters relevant to pyrolysis reactions Experiments
Modeling Challenges
Non-linear posterior and highly-correlated parameters Zero-variance data in the dataset
Application to complex pyrolysis model Conclusions
Application to pyrolysis experiments [Bessire and Minton, 2015]
Proposed five-equation model (38 unknown parameters):
R1(s) F1,1H2O + F1,2CO2+ F1,3(1-propanol)
k1
R2(s) FF2,1H2O + F2,2CO + F2,3CO2 + F2,4Phenol + F2,5Cresol + 2,6(Dimethyl Phenol) + F2,7(Trimethyl Phenol)
k2
R3(s) FF3,1CH4 + F3,2H2O + F3,3CO + F3,4H2+ F3,5CO2+ 3,6Xylene + F3,7Benzene + F3,8Toluene
k3
R4(s) k4 F4,1H2
Propagation results: production rate curves
400 600 800 1,000 1,200 1,400 0 2 4 6 ·10−6 Temperature, K Pro duction rate, s − 1 H2O H2 400 600 800 1,000 1,200 1,400 0 1 2 3 ·10−6 Temperature, K Pro duction rate, s − 1 CH4 CO 19 / 22Propagation results: production rate curves
400 600 800 1,000 1,200 1,400 0 1 2 3 ·10−7 Temperature, K Pro duction rate, s − 1 Phenol Dimethyl phenol Trimethyl phenol 400 600 800 1,000 1,200 1,400 0 1 2 3 4 ·10−7 Temperature, K Pro duction rate, s − 1 CO2 CresolPropagation results: production rate curves
400 600 800 1,000 1,200 1,400 0 2 4 ·10−8 Temperature, K Pro duction rate, s − 1 1-propanol Benzene 400 600 800 1,000 1,200 1,400 0 2 4 6 8 ·10−8 Temperature, K Pro duction rate, s − 1 Xylene Toluene 19 / 22Table of Contents
Introduction
Characterization of physico-chemical parameters relevant to pyrolysis reactions Experiments
Modeling Challenges
Non-linear posterior and highly-correlated parameters Zero-variance data in the dataset
Application to complex pyrolysis model Conclusions
Conclusions
I Application of a simple model to simulate pyrolysis experiments.
I Inference on the parameters of pyrolysis reactions using Bayesian approach.
I Efficient method to infer on kinetic parameters and characterize their uncertainties. However, lack of indentifiability for A and E .
I Rescaling the parameter space to their posterior distribution based on model features reduce significantly the tuning of the proposal covariance
I Ito-SDE method enables no proposal tuning and samples efficiently the posterior distribution, but requires the computation of gradients
I Simulations of gas production rate and mass loss with error bars lead to good agreement with experimental curves.
Future work
I Apply Itˆo-SDE method to the full set of experimental data
I Efficient computation of model gradients, e.g. adjoint-based
I Use more general models, e.g. include competitive reaction schemes (F. Torres, J. Blondeau), include more data (different heating rates)→ model inference
I Results interpretation and their used in CFD codes; uncertainty propagation for the validation of CFD codes
Acknowledgments
I This work is supported by the Fund for Research Training in Industry and Agriculture (FRIA) provided by the Belgian Fund for Scientific Research (F.R.S-FNRS)
I Scientific stay at NASA ARC was supported by the Wallonie-Bruxelles International grant WBI.World
I F. Torres (VKI), B. Helber (VKI), J. Lachaud (I2M Bordeaux), H.-W. Wong (UMassachusetts Lowell), George Bellas-C. (VKI), J.-B. Scoggins (VKI)
Bayesian parameter inference for PICA devolatilization pyrolysis at high heating rates
Joffrey Coheur
Promotors:M. Arnst1, P. Chatelain2, T. Magin4
Collaborators:P. Schrooyen3, K. Hillewaert3, A. Turchi4,
F. Panerai5, J. Meurisse5, N. N. Mansour.5 1Aerospace and Mechanical Engineering, Universit´e de Li`ege
2Institute of Mechanics, Materials and Civil Engineering, Universit´e Catholique de Louvain 3Cenaero, Gosselies
4Aeronautics and Aerospace Department, von Karman Institute for Fluid Dynamics 5NASA Ames Research Center
3rd International Conference on Uncertainty Quantification in Computational Sciences and Engineering
Additional Information
Table of Contents
Bibliography
Mathematical model Model Selection
Bibliography
I M. Bruns. Inferring and propagatin kinetic parameter uncertainty for condensed phase burning models. Fire Technology, 52:93–120, 2016.
I B. Carpenter, A. Gelman, M. D. Hoffman, D. Lee, D. Goodrich, M. Betancourt, M. Brubaker, J. Guo, P. Li, and A. Riddell. Stan: A probabilistic programming language. Journal of Statistical Software, 76(1), 2017. DOI 10.18637/jss.v076.i01
I J. Coheur, A. Turchi, P. Schrooyen and T. Magin. Development of a unified model for flow-material interaction applied to porous charring ablators. In 47th AIAA Thermophysics Conference, AIAA Aviation Forum 2017.
I N. Galagali and Y. Marzouk. Bayesian inference of chemical kinetic kinetics model from proposed reactions. Chemical Engineering Science, 123:170–190, 2015.
I H. E. Goldstein. Pyrolysis Kinetics of Nylon 6-6, Phenolic Resin, and Their Composites. Journal of Macromolecular Science: Part A - Chemistry, 3:649–673, 1969.
I B. Helber, A. Turchi, J. Scoggins, A. Hubin and T. Magin. Experimental investigation of ablation and pyrolysis processes of carbon-phenolic ablators in atmospheric entry plasmas. International Journal of Heat and Mass Transfer, 100:810–824, 2016.
I J. Lachaud and N. N. Mansour. Porous-material analysis toolbox based on Open-FOAM and applications. Journal of Thermophysics and Heat Transfer, 28(2):191–202, 2014.
I P. Schrooyen, K. Hillewaert, T. E. Magin, P. Chatelain. Fully Implicit Discontinuous Galerkin Solver to Study Surface and Volume Ablation Competition in Atmospheric Entry Flows. International Journal of Heat and Mass Transfer, 103:108–124, 2016.
I K. A. Trick, T. E. Saliba, and S. S. Sandhu. A kinetic model of the pyrolysis of phenolic resin in a carbon/phenolic composite. Carbon, 35:393–401, 1997.
I H.-W. Wong, J. Peck, R. Bonomi, J. Assif, F. Panerai, G. Reinish, J. Lachaud and N. Mansour. Quantitative determination of species production from phenol-formaldehyde resin pyrolysis. Polymer Degradation and Stability, 112:122–131, 2015
I C. Soize, Construction of probability distributions in high dimension using the maximum entropy principle: Applications to stochastic processes, random fields and random matrices, International Journal for Numerical Methods in Engineering 76 (2008).
I C. Soize, R. Ghanem, Data-driven probability concentration and sampling on manifold, Journal of Computational Physics 321 (2016) 242–258
Simple model for simulating pyrolysis experiments
For point-mass material, conduction is assumed to be instantaneous and T is uniform inside the material (lumped capacitance model, Bi = Lch
k << 1). I T = Ti = cte in ∆t = [0, tf] ξ(t) = 1− (1− n)(Ci− A exp −RTE i t) 1/(n−1) , with Ci= (1− ξ(0))1−n 1− n
I Tout = τ t, τ heating rate
ξ(T ) = 1− (1− n) −AτT exp −E RT −AτEREi −RTE + C 1−n1 C = (1− ξ0) 1− n 1−n +A τT0exp −E RT0 + Ei −E RT0 EA τ , where Ei(x)≡ Z −x −∞ exp(v ) v dv
For more complex temperature evolution, integration should be performed numerically. When Bi > 1 (Lc, h%, or k &) more complex model should be used (Argo).
Numerical set-up: 3 reactants model with 5 species
I Observable (data) dobs
ik with i ={H2,CO,CO2,CH4,H2O}: mass yields at each
temperature iteration k. Np= 3. dik = Np X j Fijm0 ξj(k)− ξj(k−1) , ∂ξj(k) ∂t = 1− ξ(k)j nj Ajexp − Ej RTk , ξj(0)= 0. 400 600 800 1000 1200 0 0.2 0.4 0.6 0.8 Temperature, T Mass yields, mg
I Simple point-mass model (0D), k = 1, . . . , nT
p={ {A3, E3, n3, FH2O,3,} R1 → H2O
{A2, E2, n2, FCO,2, FCH4,2, FH2O,2} , R2 → CO + CH4+ H2O
Posterior predictive checks: 3 reactants model with 5 species
300 400 500 600 700 800 900 1000 1100 1200 1300 0 0.2 0.4 0.6 0.8 Temperature, K Mass yields, mgPosterior predictive checks: 3 equations, 5 species
300 400 500 600 700 800 900 1000 1100 1200 70 80 90 100 Temperature, K Sample Mass, % 28 / 22How to select an appropriate model?
I Principle of Parsimony (Occam’s razor): “Shave away all that is unnecessary” Accuracy >< complexity I Kullback-Leibler information I (f , g ) = Z f (x) log f (x) g (x|θ) dx I Information criteria AIC =−2 log(L(y|ˆθ)) + 2K BIC =−2 log(L(y|ˆθ)) + K log n
Illustration with the two-equation benchmark
I Approximate model: one-equation model
400 600 800 1000 1200 1400 0 1 2 ·10−4 Temperature, K Pro duction rate 30 / 22
Illustration with the two-equation benchmark
I Approximate model: two-equation model
400 600 800 1000 1200 1400 0 1 2 ·10−4 Temperature, K Pro duction rate
Illustration with the two-equation benchmark
I Approximate model: three-equation model
400 600 800 1000 1200 1400 0 1 2 ·10−4 Temperature, K Pro duction rate 30 / 22
Illustration with the two-equation benchmark
I Computation of information criteria:
Model log-like Nparams Nsamples AIC AICc BIC
1 reaction −851.59 4 101 1711.172 1711.588 1711.188
2 reactions −46.08 8 101 108.1501 109.7153 108.1847
3 reactions −46.51 12 101 117.0258 120.5712 117.1078 3 reactions (2) −46.27 12 101 116.54 120.0855 116.592