MACHINE LEARNING FOR CELLULOSE MEMBRANES: A REPRESENTATION–MODEL–TASK FRAMEWORK FOR SMALL-DATA REGIMES
- 1V. I. Vernadsky Institute of General and Inorganic Chemistry, National Academy of Sciences of Ukraine
Abstract
Cellulose membranes are a small-data problem with a chemically ambiguous object. A polymorphic matrix, three chemically inequivalent hydroxyls at C2, C3 and C6, and a two-stage fabrication route make representation — not model architecture — the binding constraint on machine learning in this field, and this is why the toolkit that succeeded for gas-separation polymers and thin-film-composite membranes does not transfer to cellulose unchanged. This review organises the methodology around that claim. It first sets out why cellulose membranes remain of interest and where they fail, then shows that the existing computational treatment of cellulose is stratified into layers — first-principles, atomistic, coarse-grained, pore-scale, phenomenological, process-level — that do not connect, so that no chain of models runs from a synthesis recipe to a measured permeance. That specific structural gap, rather than the absence of cellulose studies in the machine-learning literature, is the argument for a data-driven approach. The representation–model–task framework is then executed task by task across property prediction, synthesis optimisation with Bayesian optimisation, inverse design, microstructural analysis and fouling forecasting, with four mandatory decisions — representation, model, task, validation — made explicit for each, and interpretability treated as a validation instrument whose attributions are falsifiable by synthesis. Cellulose/iron-oxide composites for oil–water separation serve as a running example throughout. Data resources are assessed for cellulose coverage rather than catalogued; methodological pitfalls are given cellulose-specific manifestations and mitigations; and under-explored niches are classified by operational criteria under a stated, repeatable search protocol.
Related
- Machine learning–driven inverse design of nanofiltration membranes for selective Mg2+/Na+ separation
- Valorization of cotton waste into cellulose nanocrystals and their incorporation into electrospun poly (vinyl alcohol) nanofibrous membranes with enhanced thermal and structural properties
- Insights into Desalination Performance of Spacing-Tunable Laminated MXene Membranes via Molecular Dynamics and Machine Learning
- OrbGNN: A Wave function-based Machine Learning Interelectronic Representation
- A Hybrid Message-Passing and Electric-Field Representation for Electrostatic QM/MM Embedding with Machine Learning Interatomic Potentials
BibTeX
@misc{kolomiiets2026machine,
title = {{MACHINE LEARNING FOR CELLULOSE MEMBRANES: A REPRESENTATION–MODEL–TASK FRAMEWORK FOR SMALL-DATA REGIMES}},
author = {Ye O. Kolomiiets},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009248/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009248/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - MACHINE LEARNING FOR CELLULOSE MEMBRANES: A REPRESENTATION–MODEL–TASK FRAMEWORK FOR SMALL-DATA REGIMES
AU - Ye O. Kolomiiets
PY - 2026
DA - 2026/09/22
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009248/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009248/v1
AB - Cellulose membranes are a small-data problem with a chemically ambiguous object. A polymorphic matrix, three chemically inequivalent hydroxyls at C2, C3 and C6, and a two-stage fabrication route make representation — not model architecture — the binding constraint on machine learning in this field, and this is why the toolkit that succeeded for gas-separation polymers and thin-film-composite membranes does not transfer to cellulose unchanged. This review organises the methodology around that claim. It first sets out why cellulose membranes remain of interest and where they fail, then shows that the existing computational treatment of cellulose is stratified into layers — first-principles, atomistic, coarse-grained, pore-scale, phenomenological, process-level — that do not connect, so that no chain of models runs from a synthesis recipe to a measured permeance. That specific structural gap, rather than the absence of cellulose studies in the machine-learning literature, is the argument for a data-driven approach. The representation–model–task framework is then executed task by task across property prediction, synthesis optimisation with Bayesian optimisation, inverse design, microstructural analysis and fouling forecasting, with four mandatory decisions — representation, model, task, validation — made explicit for each, and interpretability treated as a validation instrument whose attributions are falsifiable by synthesis. Cellulose/iron-oxide composites for oil–water separation serve as a running example throughout. Data resources are assessed for cellulose coverage rather than catalogued; methodological pitfalls are given cellulose-specific manifestations and mitigations; and under-explored niches are classified by operational criteria under a stated, repeatable search protocol.
ER -