# Combined Experimental and Machine Learning Study on the Interplay between Delignification and Mechanical Properties for Improved Poplar Wood Reconstruction

## TL;DR

This study improves poplar wood reconstruction by combining partial delignification with machine learning to enhance mechanical properties and sustainability.

## Contribution

A novel USS machine learning framework is introduced to predict mechanical properties based on structural and process features.

## Key findings

- Partial delignification improves mechanical properties at 45° and 90° fiber directions.
- Room-temperature delignification enables up-scaling and reuse of the solution without quality loss.
- The USS framework accurately predicts mechanical performance and identifies key fabrication parameters.

## Abstract

Structure-retaining delignification of wood is widely
used to obtain
scaffolds suitable for the preparation of high-performance biobased
composites. However, this often comes at the expense of sustainability
and large-scale production potential. To address these issues, we
reconstructed poplar wood via room-temperature partial delignification,
followed by delignification and densification. Compared to fully delignified
samples, those obtained with partial delignification have superior
mechanical properties at 45° and 90° fiber directions with
respect to the loading direction, but lower ones at 0°. Working
at room temperature facilitated sample up-scaling and allowed reuse
of the delignification solution multiple times without compromising
product quality. As shown by life cycle assessment (LCA), the possibility
of repeatedly reusing the delignification solution led to a significant
reduction in the global warming potential (GWP) and ecosystem quality
(EQ) impacts. We then developed an ’unsupervised, supervised
classification, supervised regression’ (USS) learning framework
to accurately predict the mechanical properties of reconstructed poplar
on the basis of structural and process-related features, followed
by feature importance analysis to determine the key parameters influencing
material performance. With our approach, we were able to estimate
the mechanical performance of the reconstructed samples and gain insight
into the most relevant material-fabrication parameters.