Prediction of optical spectra of coarse-grained polymers as a sequence generation problem: the Recurrent Neural Networks solution

TL;DR

This paper introduces a deep learning approach using LSTM-RNNs to predict the UV-Vis spectra of conjugated polymers directly from coarse-grained models, bypassing traditional complex back-mapping procedures.

Contribution

It presents a novel generative deep learning method that improves spectral predictions from coarse-grained polymer representations, addressing limitations of existing protocols.

Findings

Identifies discrepancies between spectra from coarse-grained models and after back-mapping.

Demonstrates the LSTM-RNN model's potential to enhance spectral prediction accuracy.

Suggests the model can improve coarse-grained potential development.

Abstract

Coarse-grained simulations of conjugated polymers have become a popular way of investigating the device physics of organic photovoltaics. While UV-Vis spectroscopy remains one of key experimental methods for the interrogation of these devices, a rigorous bridge between coarse-grained simulations and spectroscopy has never been established. Here we address this challenge by developing a method that predicts spectra of conjugated polymers directly from coarse-grained representations while avoiding ad-hoc procedures such as back-mapping from coarse-grained to atomistic representations followed by computing the spectra using standard quantum chemistry methods. Our approach is based on a generative deep learning model: the long-short-term memory recurrent neural network (LSTM-RNN) and it is suggested by the apparent similarity between natural languages and the mathematical structure of the perturbative expansions of the excited state energies due to small fluctuations of the polymer conformation. We use this model to demonstrate a dangerous discrepancy between the spectra obtained in the coarse-grained representation and after the back-mapping. This indicates that standard protocols may require additional fine-tuning in order to become reliable and that our model presents a novel tool uniquely suited for improving the back-mapping protocols and for including spectral data in the development of coarse-grained potentials.