# Estimation of the average molecular weight of microbial polyesters from FTIR spectra using artificial intelligence

## TL;DR

This paper introduces an AI-based method to estimate the molecular weight of microbial polyesters using FTIR spectra, enabling rapid and reliable analysis.

## Contribution

A novel AI method using FTIR absorbance ratios and a feature selection approach for accurate molecular weight estimation of microbial polyesters.

## Key findings

- The AI model using absorbance ratios improves robustness and reliability of molecular weight estimation.
- The method was successfully demonstrated on poly(3-hydroxybutyrate) (PHB), a microbial polyester.
- A step-by-step guide is provided to facilitate application to other polymers.

## Abstract

In this paper, we present a method for calculating the average molecular weight of microbial polyesters using Fourier transform infrared spectroscopy (FTIR) data as input. FTIR spectra provide the necessary quantitative information, as the impact of chain ends on the spectra is influenced by the average molecular weight of the polymer. Since FTIR data can be collected rapidly and is available in abundance, it serves as an ideal input for machine learning algorithms, such as artificial neural networks. The robustness and reliability of the model are improved by designing the neural network to use absorbance ratios instead of absolute absorbances as input. We also propose a new feature selection method that facilitates the identification of absorbance ratio regions best suited to serve as input for the neural network. Our approach ensures that variations in sample preparation do not compromise the accuracy of the model. The proposed computational method is demonstrated using a microbial polyester [poly(3-hydroxybutyrate), PHB], which is a biopolymer natively synthesized by multiple bacterial strains. Although the computational method has been tested with PHB, the underlying concept can be extended to other polymers. To facilitate broader application, a step-by-step guide for developing similar models is also provided.

The online version contains supplementary material available at 10.1007/s44211-025-00780-2.