Machine learning-based screening of complex molecules for polymer solar cells.

Area of Science:

• Materials Science
• Organic Electronics
• Computational Chemistry

Background:

• Polymer solar cells offer advantages like low energy payback time and scalable manufacturing.
• Current power conversion efficiencies lag behind inorganic solar cells.
• Key factors for polymer solar cell efficiency include polymer optical gap and energy level alignment with PCBM.

Purpose of the Study:

• To develop a computationally efficient screening procedure for identifying promising donor-acceptor polymers for solar cell applications.
• To reduce the cost and time associated with discovering new materials using density functional theory (DFT) calculations.

Main Methods:

• Utilized a grammar variational autoencoder model trained on a dataset of 3989 monomers derived from DFT calculations.
• Employed a simplified string representation for donor-acceptor polymers.
• The model predicts the lowest unoccupied molecular orbital (LUMO) and lowest optical transition energy without requiring atomic positions.

Main Results:

• The model accurately predicts LUMO and optical gap energies for unseen molecules with low mean absolute errors (43 meV and 74 meV, respectively).
• Demonstrated the model's capability in generating novel molecules with targeted LUMO and optical gap energies.
• Achieved a five-fold increase in the probability of finding suitable polymers compared to random search methods.

Conclusions:

• The proposed screening procedure significantly accelerates the discovery of high-performance polymer solar cell materials.
• This machine learning approach offers a cost-effective and efficient alternative to traditional computational methods for materials screening.
• The findings pave the way for the development of next-generation polymer solar cells with enhanced efficiency.