Bayesian optimal experiment design for enhanced mineral weathering with gradient boosted decision trees

[Geoengineering News]

https://www.sciencedirect.com/science/article/abs/pii/S1568494626012287

Authors: Thomas Servotte, Iris Janssens, Jakob Raymaekers, Tim Verdonck 

20 June 2026

https://doi.org/10.1016/j.asoc.2026.115780

Highlights

•Two GBDT-based batch Bayesian Opti- mization methods for mixed-variable design.

•CatBoost Ensemble outperforms GP and tree-ensemble baselines on five benchmarks.

•Applied to mineral weathering, designs beat expert-chosen ones in CDR yield.

•Framework cuts experimental cost and accelerates carbon-removal discovery.

Abstract

This paper introduces two novel methodologies for Bayesian Optimization in a batch setting, tailored to high-dimensional, mixed-variable experimental design in environmental applications, specifically carbon dioxide removal through enhanced mineral weathering. The first employs eXtreme Gradient Boosting (XGBoost) regression with the Output Configuration score to measure uncertainty, while the second uses a Categorical Boosting (CatBoost) Ensemble, deriving uncertainty from prediction variance. These methods address the challenges of complex environmental experimental spaces, offering improved scalability and adaptability over traditional Gaussian Process-based approaches. Evaluations using mixed-variable simulations on the Ackley, Griewank, Levy, Rastrigin and Schwefel benchmarks demonstrate the CatBoost Ensemble’s superior performance in high-dimensional settings. Validation on a real-world case from the Bio-Accelerated Mineral Weathering project using the CatBoost Ensemble showcases the method’s ability to outperform expert-designed experiments, accelerating the discovery of optimal conditions for carbon sequestration while minimizing resource expenditure. Our findings underscore the potential of Gradient Boosted Decision Tree-based methods, and in particular of the proposed CatBoost-based variant, to advance sustainable environmental technologies through optimized experimental design.

Source: ScienceDirect