Dielectric Polymer Property Prediction Using Recurrent Neural Networks with Optimizations.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Machine learning for predicting molecular and material properties is rapidly advancing.
• Predicting dielectric properties of polymers using computational methods is an emerging area.
• Accurate structure-property relationship prediction is crucial for materials design.

Purpose of the Study:

• To evaluate the effectiveness of recurrent neural network (RNN) models for predicting dielectric properties of polymers.
• To optimize RNN models using advanced strategies for improved accuracy and learning speed.
• To provide a C++ codebase for further research in computational materials science.

Main Methods:

• Utilized a computer-generated database of dielectric polymers.
• Implemented recurrent neural network (RNN) models, specifically Elman-type networks.
• Employed optimization strategies including binary/nonbinary SMILES fingerprints and optimized backpropagation algorithms (ATransformedBP, iRPROP- optimized).

Main Results:

• Binary SMILES representation outperformed decimal representation in training and prediction.
• Optimized RNN algorithms demonstrated superior performance over non-optimized models.
• Achieved high accuracy with average relative standard deviation (RSD) below 5% and maximum RSD not exceeding 30%.

Conclusions:

• RNN models, particularly when optimized, are effective for predicting polymer dielectric properties.
• Binary SMILES representation and advanced backpropagation are key to successful implementation.
• The developed C++ codebase supports open programming for diverse computer architectures in materials science.