Kriging and Radial Basis Function Surrogate Models for a Fluidized-Bed Methane Pyrolysis Reactor

Robust multi-objective optimization of mechanistic fluidized-bed reactor models is computationally prohibitive because each candidate design must be evaluated many times to estimate the mean and variance of its performance. This paper develops and validates computationally efficient surrogate models using Kriging (Gaussian process regression with a Matern 5/2 kernel) and radial basis function (RBF) interpolation to approximate the high-fidelity fluidized-bed methane pyrolysis reactor model. A five-dimensional design space is sampled by Latin hypercube sampling with 300 training points and 50 independent test points. At each design point, 500 Monte Carlo samples are evaluated, giving 150,000 full model runs for training and 25,000 for testing. Six primary statistics (mean and standard deviation of methane conversion, H2 mole fraction, and LCOH) are approximated. All six Kriging surrogates met the acceptance criteria (R2 ≥ 0.9956, NRMSE ≤ 1.41%, NMAE ≤ 0.69%) and outperformed the RBF benchmark for every output. The surrogates evaluate a design about 4 ? 105 times faster than Monte Carlo with an adaptive ODE solver, reducing a robust optimization run from about 3.6 days to seconds. Validation against the full model was essential: retraining with 20 additional points reduced the maximum error on the robust front from 19% to below 4.5%.