Evolutionary Approach to Constructing a Deep Feedforward Neural Network for Prediction of Electronic Coupling...

Area of Science:

• Computational materials science
• Machine learning applications in chemistry
• Electronic properties of disordered materials

Background:

• Predicting electronic coupling elements is crucial for understanding charge transport in molecular materials.
• Disordered molecular materials present significant challenges for accurate electronic property simulations.
• Current ab initio methods are computationally expensive for large-scale systems.

Purpose of the Study:

• To develop a general framework for a deep feedforward neural network (FFNN) to predict electronic coupling.
• To automate the selection of optimal FFNN architecture using an evolutionary algorithm.
• To incorporate physical properties like carrier mobility for enhanced model training.

Main Methods:

• Utilized a deep feedforward neural network (FFNN) trained on data from multiscale ab initio simulations.
• Employed an evolutionary algorithm for automated FFNN architecture optimization.
• Used Coulomb matrix representation for molecular coordinate encoding, ensuring rotation and translation invariance.
• Incorporated simultaneous minimization of model error and maximization of model fitness based on physical properties.

Main Results:

• The FFNN accurately predicted distance and orientation-dependent electronic coupling elements.
• The model achieved excellent agreement with reference data for hole transport in amorphous tris(8-hydroxyquinolinato)aluminum.
• Predicted carrier mobilities were in strong agreement with ab initio simulation results.
• The FFNN demonstrated effectiveness for transport models with and without energetic disorder.

Conclusions:

• The developed FFNN provides a powerful surrogate model for predicting electronic coupling elements and carrier mobility.
• This approach significantly overcomes the size and computational limitations of traditional ab initio methods.
• The FFNN is readily applicable to larger molecular systems with negligible computational cost.