Multiscale Digital Twin of the Physical Behavior of a Bio-Sourced Artificial Fiber Using Machine-Learned Interatomic Potentials: A First Attempt

Over the last decade, the rising demand for high-performance construction mate-rials has increasingly intersected with growing environmental concerns. This convergence highlights an urgent need for full decarbonization and the develop-ment of sustainable material solutions. As part of a strategy to reduce the carbon footprint of buildings, bio-based natural fibers, in particular cellulose-based ones, have been introduced as alternatives to conventional synthetic fibers. However, cellulose-based materials have not yet matched the mechanical performance of traditional fibers used in technical applications. A significant scientific challenge remains the limited knowledge of these fibers at the atomic scale. In this paper, a multiscale approach is adopted, utilizing molecular dynamics (MD) analysis via the LAMMPS software and a novel machine learning-based force field, MACE. Two distinct MACE models, trained on mixed and purely organic datasets, were employed to study three different scales: 1-, 3-, and 7-chain cellulose structures. As an initial exploration of this approach, the study focuses on the minimization phase to evaluate the accuracy, applicability, and computational efficiency of these machine-learned interatomic potentials.