Machine Learning and Homopolymer-Based Calibration for Enhanced Composition Analysis by Infrared Detection in Polyolefin Chromatography

Abstract

Polyolefins, the most widely produced class of synthetic polymers, play a pivotal role in modern society. Their performances are inherently controlled by key molecular parameters such as molecular weight distribution (MWD) and comonomer content—also referred to as short-chain branching (SCB)—and its distribution (chemical composition distribution, CCD, or SCBD). State-of-the-art techniques employ chromatographic methods, e.g., high-temperature gel permeation chromatography (HT-GPC/SEC), crystallization elution fractionation (CEF), and Crystallization+Extractables (CRYSTEX) in 1,2,4-trichlorobenzene (TCB) coupled with filter-based infrared (IR) detectors, providing MWD and CCD quantification. However, this approach heavily relies on labor- and resource-intensive comonomer compositional calibration workflows, requiring bespoke well-defined copolymer references/standards, narrow fractions derived from preparative molecular weight fractionation (PMMF), and 13C nuclear magnetic resonance (NMR) validation. These limitations hinder cross-laboratory consistency in data comparison and industry-wide efforts to precisely correlate molecular structures with material performance and to accelerate the development of cost-effective upcycling of plastic wastes. This study introduces a calibration strategy using commercially available α-olefin homopolymers, e.g., homopolymer polyethylene (hPE) and homopolymer polypropylene (hPP). It features three key elements: (1) spectral deconvolution of Fourier transform infrared spectroscopy (FTIR) vibration modes of CH3 and CH2 groups of hPE and copolymer solutions in TCB at 150 °C, leveraging modified Pearson VII distribution models to resolve overlapping FTIR bands of CH2 and CH3 at 2800−3000 cm−1, thus enabling the calculation of the unit absorbance for the five individual vibration modes of CH2 and CH3 groups in the wavenumbers of a filter-based IR detector; (2) experimental verification of comonomer-type-independent calibration under Beer−Lambert compliance, specifically for dilute solutions (1−4 mg/mL) with high signal-to-noise (S/N) filter-based IR detectors (IR5/IR6) coupled to HT-SEC and CEF; and (3) a dual-wavenumber machine learning (ML) approach, where the algorithms trained on a set of hPP/hPE blends correlate spectral features of the low S/N filter-based IR4 detector with stoichiometric composition. This integrated framework reduces the dependency on copolymer standards and expensive NMR instruments, offering a reliable solution for high-throughput polyolefin molecular structural characterization. Concurrently, ML models extend robust analysis to nonideal scenarios, including high-concentration CRYSTEX analysis (10 mg/mL). Validated across HT-GPC, CEF, and CRYSTEX platforms, this dual-strategy approach provides a practical pathway toward improved repeatability and reproducibility.