Machine Learning methodologies for accelerated and predictive atomic scale simulations of the transformation of biosourced molecules

Transformations of biosourced molecules are a key player in the current energy transition. The MAMABIO project aims at proposing accelerated numerical methodologies for building kinetic models of high predictive value, to develop efficient biomass transformation processes. We chose butanol dehydration reactions catalyzed by zeolites and alumina as a preferred case study. The project is organized in 3 tasks:

Task 1. Machine Learning (ML)-accelerated accurate rate constants calculations from first principles. Advanced ML tools are currently developed and will be combined to: i) calculate the energy of the targeted complex chemical system quickly and accurately, thanks to machine learning potentials derived from high-level random phase approximations (RPA) ab initio calculations, making use of machine learning perturbation theory; ii) identify most relevant collective variables that are needed to compute free energy profiles, estimate the most probable transition pathway, and calculate accurate rate constants thanks to the adaptive multilevel splitting (AMS) algorithm.

Task 2. Transient kinetic data from operando spectroscopy and chemometrics. Experimental data are gathered thanks to operando infrared spectroscopy. Chemometric analysis of the time-resolved spectra is currently being performed. Transient conditions will allow disentangling and identifying them as active, spectator or poison species. Then, quantitative thermokinetic parameters will be assessed by Multivariate Curve Resolution analysis of the spectra recorded in transient conditions. Results obtained so far are highly promising.