(387b) Machine Learning for the Study of Complex Surface Chemistries
AIChE Annual Meeting
2021
2021 Annual Meeting
Catalysis and Reaction Engineering Division
New Developments in Computational Catalysis III: Structure-Property Relationships
Tuesday, November 9, 2021 - 3:50pm to 4:10pm
We studied the performance of linear scaling versus nonlinear machine learning models for predicting adsorption and transition state energies required for the design of metal catalysts for the hydrodeoxygenation (HDO) of organic acids. Linear scaling relations to hold well for predictions of adsorption energies of surface species across metals with a mean absolute error of 0.12 eV for the test set. Here, ML methods cannot outperform linear scaling relations when the training dataset contains a complete set of energies for all species on various metal surfaces. When the training
dataset is incomplete, i.e., it only includes a random subset of species’ energies for each metal, kernel-based ML models significantly outperform linear scaling relations. Simple coordinate-free species descriptors, such as fingerprints, achieve as good results as sophisticated coordinate-based descriptors. When using the right combination of descriptors, no statistically significant difference is observed between linear and nonlinear models for predicting transition state energies. Here, bound count information is an essential descriptor together with reactant/product energies and metal descriptors. When there were missing data for reaction steps on all metal surfaces, conventional linear scaling is inferior to ML models. Next, using a stacked-GP modeling approach and an iterative approach for systematically improving adsorption and transition state energies, we predict the turnover frequencies and selectivity of various metal catalyst systems for the HDO of organic acids. Interestingly, conventional transition state scaling is not sufficiently accurate for such an iterative workflow. Only ML models, such as stacked-GP with optimized descriptors for predicting energies, are sufficiently reliable for this task. Finally, we develop a 2-dimensional volcano curve in activity and selectivity to identify optimal catalysts for the HDO of propanoic acid over metal catalysts.
dataset is incomplete, i.e., it only includes a random subset of species’ energies for each metal, kernel-based ML models significantly outperform linear scaling relations. Simple coordinate-free species descriptors, such as fingerprints, achieve as good results as sophisticated coordinate-based descriptors. When using the right combination of descriptors, no statistically significant difference is observed between linear and nonlinear models for predicting transition state energies. Here, bound count information is an essential descriptor together with reactant/product energies and metal descriptors. When there were missing data for reaction steps on all metal surfaces, conventional linear scaling is inferior to ML models. Next, using a stacked-GP modeling approach and an iterative approach for systematically improving adsorption and transition state energies, we predict the turnover frequencies and selectivity of various metal catalyst systems for the HDO of organic acids. Interestingly, conventional transition state scaling is not sufficiently accurate for such an iterative workflow. Only ML models, such as stacked-GP with optimized descriptors for predicting energies, are sufficiently reliable for this task. Finally, we develop a 2-dimensional volcano curve in activity and selectivity to identify optimal catalysts for the HDO of propanoic acid over metal catalysts.