Linking mechanics and chemistry: machine learning for yield prediction in NaBH4 mechanochemical regeneration

Abstract

Mechanochemical synthesis faces reproducibility and scale-up challenges due to complex parameter interactions. This study employs machine learning (ML) to predict NaBH₄ regeneration yield, integrating chemical experimental data and DEM (Discrete Element Method) derived invariant mechanical descriptors (Ē_n, Ē_t, f_col/n_ball). Various algorithms were evaluated, including a two-step modeling strategy to isolate the dominant effect of milling time in our process. Results demonstrate that a two-step Gaussian Process Regression (GPR) model achieves good predictive performance (R² = 0.83), significantly outperforming single-stage models and providing valuable uncertainty estimates. Tree-based ensembles (XGBoost, RF) also benefit from the two-step approach and can enhance interpretability. This work establishes a framework for using ML to optimize mechanochemical processes, reducing experimental cost and offering a method to link mechanical milling conditions to chemical outcomes, thereby enabling predictive mechanochemistry.