Machine Learning for Time-Resolved Selectivity Analysis in Methanol-To-Olefins Reaction

The methanol-to-olefins (MTO) process, a cornerstone reaction in modern coal chemical industries, generates complex time-dependent product selectivity profiles that challenge conventional data-driven modeling. Although machine learning (ML) has emerged as a promising tool for MTO reaction analysis, existing approaches often oversimplify the temporal evolution of ethylene/propylene selectivity by treating it as scalar outputs rather than functional trajectories. In this study, we propose a unified ML framework for functional outputs to model time-resolved selectivity curves. The framework employs orthogonal basis expansions to explicitly represent the output curves, thereby reducing infinite-dimensional functional data to finite-dimensional basis coefficients. Tree-based ML models are then applied to learn the mapping from inputs to basis coefficients under two distinct loss minimization strategies, achieving test-set <i>R</i>² values of up to 0.9. Shapley additive explanations value reveals that the diameter of the largest free sphere that can diffuse along a, acid density, and crystal sizes of zeolites are the shared determinants governing the selectivities to both ethylene and propylene. Notably, the largest ring sizes of zeolites exert a pronounced influence specifically on ethylene selectivity, whereas framework density of zeolites and w8 hly space velocity modulate propylene selectivity in a distinctive manner. This ML synergy provides a transferable paradigm for function output modeling of catalytic phenomena and other similar problems.