B1793
Title: Optimal experiment design for environmental research using Bayesian optimization
Authors: Thomas Servotte - University of Antwerp (Belgium) [presenting]
Iris Janssens - IDLab (UAntwerp - IMEC) (Belgium)
Tim Verdonck - KU Leuven and UAntwerpen - imec (Belgium)
Abstract: The bio-accelerated mineral weathering (BAM!) project aims to optimize the natural CO2 sequestration process of silicate weathering by finding the best combination of minerals, organic matter, and biota. The focus is on the methodological approach to design optimal experiments within a batch-based framework, wherein the combinations for subsequent batches are intelligently determined based on insights gained from prior iterations. To accomplish this, Bayesian optimization is employed, a powerful technique for optimizing complex, multi-variable systems. The dataset comprises numerous variables, many of which are categorical, posing a unique challenge. To address this, an ensemble of gradient-boosted decision trees is used, specifically the CatBoost algorithm, as a surrogate model. This ensemble method allows for the estimation of the expected outcomes and knowledge uncertainty (as opposed to data uncertainty) associated with a given combination of minerals, organic matter, and biota. Furthermore, genetic optimization is introduced into the methodology to maximize the upper confidence bound. This optimization process facilitates the identification of the most promising combinations that will most likely lead to the greatest carbon dioxide weathering effect. A batch of combinations is determined by optimizing for different exploration/exploitation trade-offs. The methodology is elaborated and some preliminary results are reported.
