Automatic Prediction of Peak Optical Absorption Wavelengths in Molecules Using Convolutional Neural Networks

Area of Science:

• Computational chemistry and materials science
• Machine learning applications in molecular design
• Optical property prediction

Background:

• Accurate prediction of molecular optical properties is crucial for applications like solar cells and batteries.
• Traditional methods like time-dependent density functional theory (TD-DFT) are computationally expensive and have inherent errors.
• Data-driven methods offer a cost-effective alternative but depend heavily on data quality and availability.

Purpose of the Study:

• To develop a cost-effective and accurate method for predicting peak optical absorption wavelengths (λmax) using machine learning.
• To leverage a multifidelity modeling approach integrating theoretical and experimental data.
• To accelerate the design of molecules with specific optical characteristics.

Main Methods:

• Utilized deep residual convolutional neural networks (DR-CNN) for feature extraction from SMILES representations.
• Employed gradient boosting feature selection and Bayesian optimization for model refinement.
• Integrated a large dataset comprising 34,893 DFT calculations and 26,395 experimental λmax values.

Main Results:

• The DR-CNN workflow, combined with gradient boosting, accurately predicts λmax solely from molecular and solvent SMILES.
• The multifidelity approach significantly enhances prediction accuracy compared to traditional methods.
• Benchmarking demonstrates superior performance against existing state-of-the-art techniques.

Conclusions:

• Learnt representations from DR-CNN integrated with machine learning accelerate molecular design for targeted optical properties.
• This data-driven approach offers a powerful alternative to computationally intensive theoretical calculations.
• The workflow facilitates the discovery of novel materials with desired optical characteristics.