POS10-1452
Accelerating Surface Passivation Material Design for High-Efficiency Perovskite LEDs via Reinforcement Learning
Topic
S10. AI-assisted Design and Simulation of Polymers
When and Where
Oct 1, 2026
08:30 - 09:30
Room 301 (Grand Ballroom)
Session Chairs
Jinkee HONG
Ki Su KIM
Presenter(s)
Dongbeen LEE (Gwangju Institute of Science and Technology)
Co-Author(s)
Abstract
Surface passivation is a key strategy to improve the external quantum efficiency (EQE) of perovskite light-emitting diodes (PeLEDs) by suppressing non-radiative recombination and stabilizing defect states. However, passivator design remains challenging because of the vast chemical space and nonlinear relationships between molecular structures and device performance. To address this, we present a reinforcement learning (RL) framework coupled with a deep neural network (DNN) predictor for EQE-enhancing passivation molecules.
The framework uses a literature-derived dataset linking molecules to EQE outcomes. Molecules are represented as SMILES strings and encoded using a pre-trained SMILES embedding model. Molecular feature processing reduces the dimensionality of pre-trained embeddings and constructs compact inputs for prediction model, while data augmentation addresses the limited size of the dataset.
The DNN predictor serves as a reward model for RL exploration. An agent proposes SMILES strings, receives predicted rewards, and shifts sampling toward molecules expected to improve EQE. Model-estimated uncertainty prioritizes reliable candidates and reduces overestimation. Repeated RL runs with different random seeds identifying recurrent candidates. By combining RL-based exploration with uncertainty estimation and cross-run recurrence analysis, this work provides a systematic machine learning framework to accelerate the design of passivators for high-efficiency PeLEDs.
This research was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2024-00342991), by the National Research Foundation of Korea(NRF) grant funded by the Korea government (MSIT) (RS-2024-00437887), by the Digital Research Innovation Institution Program through the National Research Foundation of Korea (NRF) funded by Ministry of Science and ICT (RS-2023-00283597), by Korea Toray Science Foundation.
The framework uses a literature-derived dataset linking molecules to EQE outcomes. Molecules are represented as SMILES strings and encoded using a pre-trained SMILES embedding model. Molecular feature processing reduces the dimensionality of pre-trained embeddings and constructs compact inputs for prediction model, while data augmentation addresses the limited size of the dataset.
The DNN predictor serves as a reward model for RL exploration. An agent proposes SMILES strings, receives predicted rewards, and shifts sampling toward molecules expected to improve EQE. Model-estimated uncertainty prioritizes reliable candidates and reduces overestimation. Repeated RL runs with different random seeds identifying recurrent candidates. By combining RL-based exploration with uncertainty estimation and cross-run recurrence analysis, this work provides a systematic machine learning framework to accelerate the design of passivators for high-efficiency PeLEDs.
This research was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2024-00342991), by the National Research Foundation of Korea(NRF) grant funded by the Korea government (MSIT) (RS-2024-00437887), by the Digital Research Innovation Institution Program through the National Research Foundation of Korea (NRF) funded by Ministry of Science and ICT (RS-2023-00283597), by Korea Toray Science Foundation.
