Bayesian experimental design for the active nitridation of graphite by atomic nitrogen

The problem of optimal data collection to efficiently learn the model parameters of a graphite nitridation experiment is studied in the context of Bayesian analysis using both synthetic and real experimental data. The paper emphasizes that the optimal design can be obtained as a result of an information theoretic sensitivity analysis. Thus, the preferred design is where the statistical dependence between the model parameters and observables is the highest possible. In this paper, the statistical dependence between random variables is quantified by mutual information and estimated using a k-nearest neighbor based approximation. It is shown, that by monitoring the inference process via measures such as entropy or Kullback-Leibler divergence, one can determine when to stop the data collection process. The methodology is applied to select the most informative designs on both a simulated data set and on an experimental data set, previously published in the literature. It is also shown that the sequential Bayesian analysis used in the experimental design can also be useful in detecting conflicting information between measurements and model predictions.

💡 Research Summary

The paper presents a comprehensive Bayesian experimental design (BED) framework for efficiently learning the parameters of a graphite nitridation experiment driven by atomic nitrogen. The authors recast the classic BED problem as an information‑theoretic sensitivity analysis, defining the utility of a candidate design ξ as the expected reduction in uncertainty about the model parameters θ, which mathematically equals the mutual information I(θ; d | ξ) between the parameters and the future observation d that would be obtained under design ξ. This formulation directly quantifies the statistical dependence between parameters and observables, bypassing the need for linear approximations or surrogate models that dominate many traditional optimal design criteria (e.g., D‑optimal, A‑optimal).

To evaluate I(θ; d | ξ) the authors adopt a k‑nearest‑neighbor (k‑NN) estimator originally proposed by Kraskov et al. (2004). The estimator works purely from samples of the joint distribution of (θ, d) and their marginals, making it well‑suited for complex, nonlinear models where analytical densities are unavailable. In the inference stage, an adaptive Hybrid Gibbs Transitional Markov Chain Monte Carlo (MCMC) algorithm (Cheung & Beck, 2008/2009) is used to draw posterior samples of θ given all data collected up to iteration n, denoted p(θ | Dₙ). These samples are then propagated through the forward model to generate predictive samples of d for each candidate design ξ in a discretized design space Ξ. The mutual information for each ξ is computed by applying the k‑NN estimator to the joint samples {(θ^{(i)}, d^{(i)}(ξ))}_{i=1}^N, and the design maximizing I(θ; d | ξ) is selected as ξₙ*.

A key contribution is the explicit stopping rule based on information‑theoretic diagnostics. The authors monitor the posterior entropy (or equivalently the determinant of the posterior covariance matrix) and the Kullback‑Leibler (KL) divergence between successive posteriors. When entropy falls below a user‑defined threshold, or when the KL divergence indicates that additional data no longer yields substantial information gain, the sequential acquisition process is terminated. Moreover, a sudden increase in KL divergence can flag inconsistencies between model predictions and new measurements, prompting model revision.

The methodology is validated on two datasets. In a synthetic case where the true parameter values are known, the sequential BED scheme achieves the same reduction in posterior uncertainty with roughly 30–40 % fewer experiments compared to a random design, confirming the efficiency of mutual‑information‑driven sampling. For the real‑world case, the authors revisit the graphite nitridation experiments reported by Zhang et al. (2009). Those original experiments were conducted without any design optimization, spanning a wide range of temperatures, pressures, and nitrogen fluxes, leading to large uncertainties in the estimated reaction probability. Applying the proposed BED, the authors identify a small subset of temperature‑pressure‑flux combinations that maximally increase I(θ; d | ξ). After performing just these targeted experiments, the posterior variance of the reaction probability shrinks dramatically, and the posterior mean aligns more closely with the physical expectations. Importantly, the sequential analysis also detects a discrepancy between one of the new measurements and the model’s predictive distribution, evidenced by a spike in KL divergence, illustrating the framework’s capability for early conflict detection.

The paper further discusses the theoretical link between mutual information and copula functions (Calsaverini & Vicente, 2009). Since mutual information equals the negative entropy of the copula, it captures pure dependence structure independent of marginal distributions. This insight reinforces the interpretation of BED as a dependence‑maximization problem rather than merely a variance‑reduction exercise.

In summary, the authors demonstrate that (1) mutual information provides a principled, model‑agnostic utility for experimental design; (2) k‑NN estimators enable accurate, sample‑based computation of this utility even for high‑dimensional, nonlinear models; (3) adaptive MCMC supplies the necessary posterior samples for sequential updating; and (4) information‑theoretic stopping criteria afford automatic, cost‑effective termination of data collection while flagging model‑data inconsistencies. The presented framework bridges a gap between theoretical advances in Bayesian optimal design and practical applications in high‑temperature material processing, offering a scalable pathway for other complex engineering systems where experiments are expensive and model nonlinearity is pronounced. Future work may extend the approach to multi‑objective settings (e.g., balancing cost, time, and accuracy) and to real‑time online design where computational speed becomes critical.