Can Artificial Intelligence Design Your Experiment? A Carbon Capture Example

Abstract In this paper, we introduce a sequential design of experiments (SDOE) methodology tailored for laboratory research, with a specific focus on novel chemistry for carbon capture. By intelligently guiding the selection of experiments, SDOE iteratively enhances data model performance while reducing the number of required experiments by up to 50%, ensuring more efficient and impactful outcomes. The proposed SDOE methodology is based on modern active learning principles, offering an adaptive and iterative approach to experimentation. It dynamically adjusts upcoming experiments and accommodates major changes to the experimental program, such as introducing new features, parameters, or refining feature ranges. At the core of this methodology is a machine learning (ML) model that is automatically retrained after each experiment. Using various active learning strategies (acquisition functions), including exploration, exploitation, and hybrid approaches, SDOE identifies the next experiment with the highest expected information gain. The novel carbon capture system exhibits highly variable physico-chemical properties across different compositions and conditions. In the SDOE workflow, we selected Gaussian process regression (GPR) as a robust ML algorithm for predicting physical properties (the direct problem) and CO2 concentration (the inverse problem). Starting with a small initial training dataset, the SDOE process trains an initial model to predict responses across a pool of potential experiments. Active learning strategies then guide the selection of subsequent experiments based on various criteria such as uncertainty (for exploitation), diversity (for exploration), or hybrid methods balancing uncertainty and diversity. Each completed experiment adds new data to the training set, allowing the model to be iteratively updated. This iterative approach significantly reduces the number of experiments required to develop accurate predictive models, cutting time, cost, and resource utilization while maximizing experimental efficiency. Overall, this SDOE methodology accelerates the novel carbon capture technology optimization by integrating adaptive, AI-driven strategies into laboratory programs. Its flexibility allows for real-time adjustments and resource efficiency. Having successfully developed and demonstrated SDOE for this application, we aim to extend this active-learning-based approach to other experimental fields, promising substantial economic and scientific benefits.