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

Area of Science:

• Chemical Engineering
• Catalysis Science
• Machine Learning Applications

Background:

• The methanol-to-olefins (MTO) process is crucial in coal chemical industries but exhibits complex, time-dependent product selectivity.
• Conventional data-driven models struggle to capture the temporal evolution of ethylene/propylene selectivity, often treating it as scalar rather than functional data.

Purpose of the Study:

• To develop a unified machine learning (ML) framework for modeling time-resolved selectivity curves in MTO reactions.
• To represent product selectivity as functional trajectories rather than simplified scalar outputs.

Main Methods:

• Employed orthogonal basis expansions to convert infinite-dimensional functional data into finite-dimensional basis coefficients.
• Utilized tree-based ML models to learn the input-to-basis coefficient mapping.
• Applied two distinct loss minimization strategies for model training.

Main Results:

• Achieved high prediction accuracy with test-set R² values up to 0.9 for time-resolved selectivity curves.
• Identified key determinants of ethylene and propylene selectivity, including zeolite properties (largest free sphere diameter, acid density, crystal size) and process parameters (framework density, space velocity).
• Discovered specific influences: largest ring sizes on ethylene selectivity, and framework density/space velocity on propylene selectivity.

Conclusions:

• The proposed ML framework offers a robust approach for modeling functional outputs in catalytic reactions.
• This methodology provides a transferable paradigm for analyzing complex time-dependent phenomena in chemical processes and beyond.
• The findings enhance understanding of MTO reaction mechanisms and catalyst design.