Wider research context
Transitioning to sustainable energy sources requires replacing fossil-based fuels with synthetic energy vectors. Chemical energy storage in high-energy-density molecules catalyzed by heterogeneous catalysts is one of the most attractive routes for this aim. In situ and operando experimental techniques have shown that such materials are highly complex under reaction conditions since many physicochemical processes co-occur. Theoretical calculations are essential to comprehend them at the atomic level and provide the path to designing tailored catalytic systems. However, state-of-the-art methods fail to capture the catalysts’ intricate complexity; a vast gap exists between “real” catalysts in the lab and current theoretical approaches. Current modeling methods are limited to specific short-time/small-length scales, simple catalytic models, and static active sites, which hamper the predictive power of simulations.
Hypotheses and Objectives
The original DYNAMO hypothesis is that revisiting the “active” site concept and considering it “dynamic” can pave the way for the theory-guided design of optimized catalysts. Thus, to drive theory development, we will study complex catalytic systems relevant to the energy field: the CO2 hydrogenation to methanol catalyzed by SiO2-supported Cu-based alloyed nanoparticles. For that aim, DYNAMO will develop high dimensional potentials based on deep neural networks for these complex materials with nearly first principles accuracy.
Approach and Methods
We will train the neural networks using first-principles molecular dynamics using advanced density functional theory (DFT) methods for the electronic structure. We will use metadynamics as the initial sampling technique to explore the free energy surface (FES), including chemical reactions and active site restructuring. We will formalize the investigation of the free energy surfaces as a Markov Decision, Process (MDP) by applying a bias potential to explore the FES in the spirit of metadynamics.
Level of originality and Innovation
Unlike metadynamics, the form of the bias potential will be defined by a parametrized policy not defined beforehand and determined by a reinforcement learning (RL) algorithm. Combining the developed potentials and the RL algorithm will provide a general framework to study operando complex materials and capture their dynamic behavior. We will use cutting-edge and novel theoretical approaches in first principles molecular dynamics, deep learning, and reinforcement learning.
Primary researchers involved
Professor Aleix Comas-Vives from TU Wien will lead the proposal and supervise two co-workers funded by the project. Two international experimental collaborators with whom the PI has active collaborations: Prof. Christophe Copéret (ETH Zürich) and Senior Scientist Olga Safonova (PSI), will provide reliable experimental data to benchmark the methodologies and test the potential of the theoretical predictions.
