Machine Learning Model Predicts Residual Strength of Aged Silk Heritage Without Damage

Silk textiles in museum collections can lose most of their tensile strength while looking outwardly unchanged, and the only direct way to measure how much strength remains—pulling a sample apart on a tensile tester—is a destructive test that conservators cannot ethically apply to irreplaceable artifacts. A new study in the journal Fashion and Textiles reports a way around that impasse: a machine learning model that estimates the residual tensile strength of silk from measurements that can all be taken without damaging the fabric. The work, led by Seonghee Pak, Byoungho Ellie Jin, and Changsang Yun, combined nondestructive spectroscopy, surface color analysis, pH measurement, and moisture testing on artificially aged silk, then trained an XGBoost algorithm to translate those signals into a quantitative strength estimate.

The research team began with six types of 100% silk fabric made from Bombyx mori filament, differing in degumming status, weave type, density, yarn count, weight, and thickness. None of the fabrics was bleached or dyed. To simulate the stresses that heritage textiles face, the researchers exposed samples to three kinds of accelerated deterioration: light in a sunlit indoor environment simulated with a xenon arc lamp behind glass filters, UV-A radiation from a UV-A 340 lamp, and dry heat in an oven at temperatures of 100, 150, 175, and 200 degrees Celsius. Analyses were conducted at cumulative exposure times of 0, 24, 72, 120, and 168 hours, with extended monitoring under the slower heat conditions. One fabric, designated sample F, showed the most representative deterioration behavior and served as the primary model for presenting results, while data from all six fabrics fed the statistical analyses and the machine learning dataset.

The nondestructive measurements centered on the protein structure of silk fibroin itself. Using Fourier transform infrared spectroscopy in attenuated total reflectance mode—a technique the authors note is commonly accessible in museums—they tracked the amide I region of the spectrum, where peak positions reflect the protein’s secondary structures: α-helix, β-sheet, β-turn, and random coil. As deterioration progressed, the amide I peak shifted toward higher wavenumbers, a blue shift the researchers attribute to weakening hydrogen bonding as the carbon-oxygen bonds become relatively strengthened. Fourier self-deconvolution of the amide I band resolved eight subpeaks, allowing the team to calculate the relative content of each secondary structure and a crystallinity index based on the ratio of β-sheet to random coil absorbance.

The structural story that emerged was stage-dependent. In the early stages of deterioration, the α-helix and random coil contents—components of the loosely ordered amorphous regions—decreased, while β-sheet content increased. The researchers interpret this initial rise as a relative enrichment: the fragile amorphous regions broke down first, leaving the more robust crystalline β-sheet domains behind. The crystallinity of the untreated sample was 44.29%, rising to 45.71% to 45.98% early in the aging process. Crucially, the authors caution that this rising crystallinity does not signal improved structural integrity; the concurrent blue shift reflects weakening inter-chain hydrogen bonds and bond breakage, so tensile strength fell even as the crystalline proportion climbed. Later, β-sheet content began to decline—after only 6 hours under the harshest 200-degree heat condition, and markedly under the light and UV-A conditions—indicating that the crystalline regions themselves eventually collapse.

Tensile strength, measured destructively on the aged samples with a universal testing machine according to ISO 13934-1, declined with cumulative stress across all treatment groups. The rate depended heavily on the severity of the stress. Under the 200-degree heat condition, strength loss occurred most rapidly, with the experiment terminated when strength converged to zero. Most tensile strength was lost after 168 hours under the simulated indoor light and UV-A conditions. By contrast, the milder 100-degree and 150-degree heat conditions produced only about 40.1% and 44.2% strength loss even after 168 hours, and the reduction remained below 50% after 400 hours of exposure. The authors interpret this as dry heat deterioration responding more sensitively to temperature increases than to exposure duration. Interestingly, the type of deterioration environment did not significantly influence the strength reduction pattern; rather, strength loss tracked the magnitude of applied energy.

Alongside the structural changes, the team measured pH, moisture regain, yellow index, and K/S values. The pH of the untreated sample was 8.6, and it fell sharply in the early stages of deterioration—dropping to 6.19 under the light condition, 5.24 under UV-A, and 6.62 within 48 hours under the harshest heat condition. The researchers link this acidification to acidic products such as carboxyl groups generated during deterioration, consistent with earlier findings that main-chain hydrolysis produces acidic residues. Moisture regain, measured under museum conservation conditions of 20 degrees Celsius and 50% relative humidity, fell from 8.61% in the untreated fabric, with the fastest decrease under the harshest heat condition. Yellowing increased with cumulative exposure, most rapidly under the harshest heat, and the K/S values calculated from reflectance at 400 and 580 nanometers confirmed the visible shift: greater absorption of violet light deepened the perceived yellow tint while reducing brightness.

Correlation analysis revealed which of these indicators actually track tensile strength. Tensile strength showed statistically significant positive correlations with β-sheet and α-helix contents, and significant negative correlations with crystallinity, random coil content, β-turn content, and the amide I shift ratio. Notably, pH, moisture regain, yellow index, K/S value, and the type of stress did not show direct statistical significance with tensile strength. However, pH was closely associated with variables that were significant—crystallinity, β-sheet content, and random coil content—suggesting that some factors contribute to mechanical deterioration indirectly by mediating structural rearrangements. A linear regression model built on the significant indicators achieved an explanatory power of 57.6%, statistically significant but limited, which the authors attribute to nonlinear relationships and complex interactions among the variables that simple linear models cannot capture.

That limitation motivated the machine learning approach. The researchers assembled a dataset of 95 independent observations, each representing a unique combination of fabric type and deterioration duration, and split it into 80% training and 20% test sets. They selected XGBoost, an extreme gradient boosting algorithm, because it excels at modeling nonlinear relationships and interactions in relatively small datasets, and they tuned hyperparameters through randomized search cross-validation. The first model, using only features with high Pearson correlation coefficients, reached an R-squared of 0.403. A second model incorporating all degradation-related indicators improved to 0.654, with corresponding decreases in root mean square error, mean absolute error, and mean absolute percentage error. The best performance came from a model designed for undyed textile heritage, which used crystallinity, protein secondary structure contents, pH, yellow index, and K/S value and achieved an R-squared of 0.751—the highest accuracy of any configuration tested.

A fourth model addressed a practical constraint of real collections: many textile artifacts are too small to permit reliable moisture regain measurement, and dyed fabrics complicate yellow index readings. By excluding the yellow index and K/S value, this configuration offers potential utility for predicting the residual strength of dyed fabrics under constrained conditions. The machine learning results also demonstrated that predictive accuracy depends on the combination of variables rather than any single dominant indicator—even variables with modest individual standardized coefficients contributed meaningfully when paired with others. The authors acknowledge that the models’ predictive performance, spanning R-squared values of 0.403 to 0.751, may not yet be sufficient to fully replace direct mechanical testing, but they argue the models can serve as useful references for risk assessment and conservation planning, where quantitative criteria for evaluating mechanical strength remain scarce.

The study’s scope comes with clear boundaries. All samples were undyed silk fabrics artificially aged under controlled, reproducible laboratory conditions, whereas the aging environments of real cultural heritage—such as materials recovered during excavation—are inherently complex and highly variable. The authors note that future work must address dye fading, extend the methodology to other natural fibers such as wool and ramie and to man-made fibers, refine the quantitative models, and develop practical guidelines for allowable loads. They also emphasize that accumulating data on textiles with diverse properties and expanding the indicator pool will enhance nondestructive predictive performance. Even so, the study demonstrates a meaningful step toward overcoming the limitations of subjective condition assessment, in which skilled conservators evaluate damage based on professional experience recorded in subjective language that is difficult to share or compare. A quantitative, nondestructive estimate of residual tensile strength could give conservators, curators, and administrators a common language for decisions about long-term display planning, international loan exhibitions, and safe handling of garments, socks, pouches, and cushions that experience physical stress every time they are moved, exhibited, or treated.

Subject of Research: Nondestructive machine learning prediction of residual tensile strength in artificially aged silk textile cultural heritage

Article Title: Nondestructive quantitative model for predicting the residual tensile strength of silk textile cultural heritage based on machine learning

Article References: Pak, S., Jin, B. E., & Yun, C. (2026). Nondestructive quantitative model for predicting the residual tensile strength of silk textile cultural heritage based on machine learning. Fashion and Textiles, 13(1), Article 23. https://doi.org/10.1186/s40691-026-00472-z

Image Credits: AI Generated

DOI: 10.1186/s40691-026-00472-z

Keywords: silk, textile conservation, machine learning, XGBoost, tensile strength, FTIR spectroscopy, crystallinity, protein secondary structure, cultural heritage, nondestructive analysis, accelerated aging, pH