A multiscale Machine Learning based Software for the Simulation of Catalytic Processes

The reduction of the environmental footprint of the chemical and related industries is nowadays of utmost importance.

The transition towards more sustainable processes that combine efficient use of raw material and energy with higher transformation rates, better selectivity and higher mass and energy efficiency will contribute to meet the objectives of the green deal.

In this respect, catalysis engineering is pivotal to developing technologies able to meet these goals and to shape the sustainable economy of the future.

The accurate description of this multiscale process has a substantial impact on the performances of the entire chemical process and, consequently, on many manufacturing sectors.

The description of the catalytic process requires a detailed and accurate definition of the intrinsic reactivity, by means of first-principles kinetic schemes, coupled with rigorous models at the reactor scale.

Currently, this approach is hindered by the limited available computational resources which prevent the adoption of detailed and atom-resolved kinetic models into reactor simulations with a reasonable computational burden.

To overcome the limitations identified above, starting from the results obtained during the ERC Stg SHAPE (n. 677423), we propose MultiCAT, a highly accurate yet computationally lean multi-scale physics-guided machine learning-based surrogate modelling framework of the entire reactor from the atomistic to the process scales.

This represents a leapfrog improvement in the detailed numerical modeling of catalytic processes, by achieving a drastic reduction in the computational cost with a concomitant boost in the prediction reliability, and paving the way for a new generation of catalytic process models, an evolution of hybrid digital twins, for online process design, optimization and control.