Book Scientists Use Microscopy, Imaging and AI to Preserve Texts

Specialists working at the intersection of conservation science and data science are increasingly using microscopes, multispectral imaging and machine learning to read, analyze and preserve fragile books and manuscripts. In a first-person account published April 5, 2026, PhD student Christina Dinh Nguyen (University of Toronto) describes collaborative projects that pair imaging-based analysis with conservation practice to study parchment manuscripts and materials. Parallel research and publicized projects have demonstrated machine-learning models decoding highly damaged artifacts — from burnt Roman scrolls to crumbling cuneiform — and automated handwritten-text recognition (HTR) systems applied across medieval collections.

Context Cultural heritage collections worldwide face accelerating threats: climate-driven humidity and temperature shifts, physical degradation of organic supports such as parchment and papyrus, and destruction from conflict and disasters. At the same time, large-scale digitization efforts have made high-resolution images available, opening opportunities for computational analysis. As digitized images proliferate, AI-driven methods can convert visual data into searchable, machine-readable text and extract material information that informs conservation choices.

Researchers deploy a suite of tools. Optical microscopy and material analysis reveal fibres, ink composition and surface degradation patterns that guide physical interventions. Multispectral and hyperspectral imaging capture reflectance beyond the visible range, enhancing the contrast between ink and substrate and revealing erased or faded text. On the computational side, supervised and deep-learning models are trained on labeled images to perform handwriting recognition, ink-detection and virtual unwrapping or segmentation of layered documents. Trinity College Dublin has summarized advances in HTR that translate image sequences into transcripts across varied scripts and languages, while high-profile cases reported in scientific outlets have shown neural networks recovering text from charred or tightly rolled scrolls.

These techniques are already producing concrete results: teams have used machine learning to detect carbon-based ink on Herculaneum scrolls and other damaged manuscripts, and neural networks have helped extract information from cuneiform fragments and scorched Roman texts. Institutions are pairing imaging outputs with conservation workflows so that non-invasive scanning can inform treatment plans and digitized transcriptions increase accessibility for scholars and the public.

Implications The combination of imaging and AI is shifting what is possible for manuscript studies and conservation. Machine-readable outputs accelerate scholarship by enabling search, corpus-level analysis and cross-collection comparisons; material insights from imaging inform prioritized interventions and environmental controls. However, technical and ethical considerations remain: HTR and model generalization vary across scripts and languages, digitization quality and metadata matter for algorithmic performance, and stewardship decisions must weigh access against risks to fragile objects.

Follow-up developments to watch include broader adoption of standardized imaging protocols and interoperable metadata to improve model transferability, publications reporting benchmarks for HTR on non-Latin scripts, and collaborative projects that integrate conservation outcomes with computational results. Ongoing high-profile recoveries (for example, work on Herculaneum scrolls and other damaged collections) will continue to test and publicize these approaches.