CHEMFLUOR-VAE: REVERSE DESIGN OF ORGANIC FLUOROPHORES BASED ON EXPERIMENTAL OPTICAL PROPERTIES AND VARIATIONAL AUTOENCODER

xu, yuzhi

Permalink

https://hdl.handle.net/2142/122688 Copy

Description

Title

CHEMFLUOR-VAE: REVERSE DESIGN OF ORGANIC FLUOROPHORES BASED ON EXPERIMENTAL OPTICAL PROPERTIES AND VARIATIONAL AUTOENCODER

Author(s)

xu, yuzhi

Contributor(s)

• Zhang, Haozhe
• Mi, Zehan
• Lin, Ruiming
• Bai, Hanzhi
• Li, Bo
• Luo, Yongrui
• Ju, Chengwei

Issue Date

2023-06-23

Keyword(s)

Theory and Computation

Abstract

Organic fluorescent molecules with desired optical properties attracted great attention, while the rational design was hindered by unclear structure properties relationship and the lack of rapid/affordable prediction methods. With the introduction of statistics-based methods in the prediction of photophysical properties for organic dyes, reverse design of fluorophores without traversing chemical space is still challenged by the features used for current methodologies. In this work, we construct a self-referencing embedded strings (SELFIES)-based variational autoencoder (VAE) and a prediction model, which uses the latent space as the input, for the organic fluorophores, in the absence of joint training. The VAE can reproduce the structure of midsize organic dyes with acceptable accuracy. A tree-based prediction model based on Gradient Boosted Regression Trees (GBRT) can estimate the optical properties of organic dyes with a MAE 0.134 eV for emission energy and an accuracy of 0.81 for photoluminescence quantum yield (PLQY), which is comparable with the state-of-the-art quantum-mechanical based approach, time-dependent density-functional theory (TD-DFT). The feasibility of our approach in reverse design is proved by preliminary attempts at skeleton optimization and validated by first-principles calculations. New experimental synthesized molecules demonstrated the accuracy of our prediction model. Meanwhile, due to the continuous values in the latent space, this VAE-based methodology makes gradient optimization become possible for large organic materials. Combined, our statistical learning methodology opens a new venue for the design of organic fluorophore, can also be extended to the field of organic solar cell (photo conversion efficiency, PCE) and organic field-effect transistor (conductivity).

Publisher

International Symposium on Molecular Spectroscopy

Type of Resource

Text

Language

eng

Handle URL

https://hdl.handle.net/2142/122688

Owning Collections