Learning design principles for photoinduced molecular processes from ML-enhanced quantum dynamics simulations
||Learning design principles for photoinduced molecular processes from ML-enhanced quantum dynamics simulations|
||Nanna Holmgaard List <email@example.com>|
||Kungliga Tekniska högskolan|
||2023-02-15 – 2023-09-01|
Light–matter interactions underpin an ever-expanding range of technological applications, such as photovoltaics, sensors, imaging and photocatalysts. Photons offer multiple
advantages as low-loss energy carriers but also in the context of matter: providing a window into its quantum nature, triggering and driving transformations as well as externally controlling them. Critically, the additional energy introduced by light can drive processes in molecules and materials significantly different from those available under thermal conditions. However, a key roadblock to harnessing this capacity is that we do not understand the evolution of the photoinitiated nonequilibrium state, let alone how to control it. Yet, there is an urgent need to enable optimization and design of molecules and molecular materials with targeted photoinduced functions.
Simulations rooted in quantum mechanics, such as ab Initio nonadiabatic dynamics simulations, are uniquely poised to decipher such processes by providing direct access to the underlying coupled electronic and nuclear structure. However, there are two pressing bottlenecks to address. First, the high computational cost of molecular quantum simulations severely restricts the time scales and level of system complexity that can be addressed. This becomes particularly problematic when considering processes involving several spin manifolds. Second, deriving insights from the wealth of coupled electronic and nuclear data produced by such simulations toward formulating design principles remains a major challenge. Machine learning techniques have the potential to overcome these challenges and significantly contribute to advancing this field.
In this project, we will leverage first-principles molecular quantum dynamics simulations together with machine-learning frameworks (e.g., PyTorch and TensorFlow) to train ML models with the aim of: (i) predicting excited-state quantities in order to enable ML-accelerated nonadiabatic dynamics simulations and (ii) automatically extracting essential insights from the coupled electronic and nuclear data about the underlying rules governing photoinduced behaviors.
This project is motivated by our recent work* on combining physics-based descriptors and excited-state data to train ML classification models to predict photo physical properties: *“Dimming the lights in GFP: High-throughput, in silico mutagenesis and machine learning predict the fluorescence of green fluorescent protein variants”, Jones, List et al. (to be submitted, 2023).