Simulation-based optimal Bayesian experimental design for nonlinear systems

The optimal selection of experimental conditions is essential to maximizing the value of data for inference and prediction, particularly in situations where experiments are time-consuming and expensive to conduct. We propose a general mathematical framework and an algorithmic approach for optimal experimental design with nonlinear simulation-based models; in particular, we focus on finding sets of experiments that provide the most information about targeted sets of parameters. Our framework employs a Bayesian statistical setting, which provides a foundation for inference from noisy, indirect, and incomplete data, and a natural mechanism for incorporating heterogeneous sources of information. An objective function is constructed from information theoretic measures, reflecting expected information gain from proposed combinations of experiments. Polynomial chaos approximations and a two-stage Monte Carlo sampling method are used to evaluate the expected information gain. Stochastic approximation algorithms are then used to make optimization feasible in computationally intensive and high-dimensional settings. These algorithms are demonstrated on model problems and on nonlinear parameter estimation problems arising in detailed combustion kinetics.

💡 Research Summary

The paper presents a comprehensive framework for optimal experimental design (OED) in the context of nonlinear, simulation‑based models, using a fully Bayesian approach. The authors adopt expected information gain (EIG), measured as the Kullback‑Leibler divergence from the posterior to the prior, as the design utility. This choice aligns with Lindley’s decision‑theoretic formulation and is equivalent to the mutual information between model parameters and potential data.

Direct evaluation of EIG requires integration over both the parameter space and the data space, which is computationally prohibitive for expensive forward models. To overcome this, the authors construct a generalized polynomial chaos (PC) surrogate that approximates the model output as a high‑order polynomial function of both parameters and design variables. The surrogate is built using dimension‑adaptive sparse quadrature, automatically detecting anisotropic dependencies and reducing the number of required model evaluations.

EIG estimation is performed with a two‑stage Monte‑Carlo (MC) scheme. In the first stage, samples are drawn from the prior distribution of the parameters; in the second stage, for each parameter sample, design‑space samples are generated and the corresponding model outputs are evaluated via the PC surrogate. This yields an unbiased estimator of the double integral defining EIG while dramatically lowering the computational cost. Variance‑reduction techniques such as importance sampling and control variates are incorporated to improve estimator efficiency.

Optimization of the design variables is carried out using stochastic approximation (SA) algorithms, specifically Robbins‑Monro and Kiefer‑Wolfowitz type updates. Because the EIG estimator is noisy, SA provides a principled way to update the design iteratively using noisy gradient estimates, guaranteeing convergence under standard step‑size conditions even in high‑dimensional design spaces. The framework naturally extends to batch designs where multiple experiments are planned simultaneously; the design vector is expanded to an N‑fold dimension and the joint likelihood of all experiments is used in the EIG formulation. The authors prove that the expected utility of a batch is not simply the sum of individual utilities, highlighting the necessity of joint optimization.

Two illustrative examples validate the methodology. The first is a simple nonlinear two‑parameter model where the proposed approach identifies the optimal design with roughly ten times fewer forward model calls than exhaustive grid search, achieving identical utility values. The second example involves a detailed combustion kinetics problem with hundreds of reaction parameters. Here, the PC surrogate replaces on the order of 10⁴ expensive CFD simulations, and the optimal batch of five ignition experiments reduces posterior parameter uncertainty by about 40 % and improves predictive error by roughly 30 %.

Key contributions of the work are: (1) a rigorous Bayesian OED criterion based on expected Shannon information gain for nonlinear models; (2) an efficient surrogate‑based estimator for the double integral defining EIG; (3) a scalable stochastic optimization scheme that handles continuous and batch design spaces; and (4) demonstration on a realistic high‑dimensional scientific application. By integrating information‑theoretic design, surrogate modeling, and stochastic optimization, the paper advances the state of the art beyond linear‑Gaussian approximations, locally optimal criteria, and brute‑force enumeration, offering a practical tool for scientists and engineers facing costly experiments and complex models.