[Paper Review] Bayesian Design of Experiments: Implementation, Validation and Application to Chemical Kinetics

Eric A. Walker, Kishore Ravisankar|arXiv (Cornell University)|Sep 9, 2019

Advanced Multi-Objective Optimization Algorithms46 references4 citations

TL;DR

This paper presents a Bayesian experimental design (BED) framework for optimizing experiments in chemical kinetics, using expected information gain to guide choices in experimental variables such as temperature and reactor volume. It demonstrates the method through three case studies—linear regression, a reaction free energy model, and a catalytic membrane reactor—validating the approach with numerical and analytical benchmarks, and optimizing design via differential evolution and steepest ascent algorithms with grid search initialization.

ABSTRACT

Bayesian experimental design (BED) is a tool for guiding experiments founded on the principle of expected information gain. I.e., which experiment design will inform the most about the model can be predicted before experiments in a laboratory are conducted. BED is also useful when specific physical questions arise from the model which are answered from certain experiments but not from other experiments. BED can take two forms, and these two forms are expressed in three example models in this work. The first example takes the form of a Bayesian linear regression, but also this example is a benchmark for checking numerical and analytical solutions. One of two parameters is an estimator of the synthetic experimental data, and the BED task is choosing among which of the two parameters to inform (limited experimental observability). The second example is a chemical reaction model with a parameter space of informed reaction free energy and temperature. The temperature is an independent experimental design variable explored for information gain. The second and third examples are of the form of adjusting an independent variable in the experimental setup. The third example is a catalytic membrane reactor similar to a plug-flow reactor. For this example, a grid search over the independent variables, temperature and volume, for the greatest information gain is conducted. Also, maximum information gain is conducted is optimized with two algorithms: the differential evolution algorithm and steepest ascent, both of which benefitted in terms of initial guess from the grid search.

Motivation & Objective

• To develop and validate a Bayesian experimental design (BED) framework for chemical kinetics that maximizes information gain before conducting experiments.
• To address limited experimental observability by selecting which parameters to measure, using expected information gain as the optimization criterion.
• To apply BED to real chemical systems, including a reaction free energy model and a catalytic membrane reactor, to guide optimal experimental setup.
• To compare and validate numerical and analytical solutions in a benchmark linear regression model to ensure methodological accuracy.
• To optimize experimental design variables (e.g., temperature, volume) using global and local optimization algorithms, informed by grid search initialization.

Proposed method

• Formulates BED using expected information gain as the objective function, quantifying how much each experimental design reduces uncertainty in model parameters.
• Applies Bayesian linear regression as a benchmark model to validate numerical solutions against analytical results.
• Uses a chemical reaction model with free energy and temperature as parameters, treating temperature as an adjustable experimental design variable.
• Implements a grid search over temperature and reactor volume to identify promising regions for maximum information gain.
• Employs differential evolution and steepest ascent algorithms to optimize the experimental design, with initial guesses informed by the grid search.
• Validates results through comparison with analytical solutions in the linear regression case and simulation-based assessment in the chemical kinetics models.

Experimental results

Research questions

• RQ1 Which experimental design yields the highest expected information gain for estimating unknown parameters in a Bayesian framework?
• RQ2 How can limited experimental observability be addressed by selecting which parameters to measure based on information gain?
• RQ3 To what extent can numerical optimization methods like differential evolution and steepest ascent improve experimental design in chemical kinetics?
• RQ4 How does grid search initialization enhance the convergence and performance of optimization algorithms in BED?
• RQ5 Can the BED framework be reliably validated using analytical benchmarks and applied to realistic chemical systems?

Key findings

• The Bayesian experimental design framework successfully identifies optimal experimental conditions before physical experimentation, minimizing resource waste.
• The benchmark linear regression model confirmed the accuracy of numerical solutions by matching them closely with analytical results.
• In the reaction free energy model, temperature was identified as a key design variable that significantly influences information gain.
• For the catalytic membrane reactor, optimal designs were found at specific combinations of temperature and volume, with information gain maximized through optimization.
• Both differential evolution and steepest ascent algorithms effectively located high-information designs, with the grid search providing robust initial guesses.
• The method demonstrated scalability and robustness across diverse chemical kinetics problems, supporting its use in real-world experimental planning.