Herculaneum papyri: reading without unrolling

Sun, 04 Oct 2026 00:00:00 +0000

Fragments of papyrus PHerc. 1521, held by the British Library. Photographer unknown, British Library; source page, public domain.

The Herculaneum scrolls look like lumps of charcoal. Yet Greek writing remains inside those dark cylinders. The eruption of Vesuvius in AD 79 buried the town and the Villa of the Papyri, where an exceptional ancient library was found. The heat carbonized the scrolls: that transformation helped them survive, but also made them brittle and difficult to open.

The question is one of engineering and conservation at once: how can we recover writing from an object that might be destroyed if we try to unfold it? The answer combines experimental physics, three-dimensional reconstruction, image processing, machine learning and expert reading. None of these stages is enough on its own.

Herculaneum before and after the eruption

Herculaneum was a compact coastal town about seven kilometres from Vesuvius, with decorated houses, baths and public buildings. The Archaeological Park estimates around twenty hectares and four thousand residents. The town still bore the marks of the AD 62 earthquake when Vesuvius changed its landscape for good in AD 79.

AI-generated illustration imagining the town before and after the AD 79 eruption. Building positions, terrain and coastline are illustrative; this is not an exact archaeological reconstruction.

The town stood on a terrace above the sea, between two streams. Its streets formed a grid, and public buildings such as the theatre and basilica had been refurbished under Augustus. Unlike the familiar picture of Pompeii, Herculaneum also preserved doors, furniture, food and other organic material. These extraordinary remains help reconstruct daily life; the same event that preserved them made the villa’s books almost impossible to read. The Archaeological Park describes these finds.

The eruption sent up ash and pumice; hours later, collapsing parts of the column drove dense clouds of hot gas and ash across Herculaneum to the shoreline. The Park gives temperatures of around 400 °C. A 2023 geological study estimated 495–555 °C for the first surge from charcoal in wood, with cooler later currents. These are reconstructions, not a single temperature measured across the town; duration, air and water mixing, and method affect the estimates. Deposits about twenty metres deep buried Herculaneum.

Remains found in the waterfront chambers, known as the fornici. Photo by Ad Meskens, Wikimedia Commons; original page, CC BY-SA 4.0. Resized and converted to WebP.

On the beach and in twelve stone chambers by the water, archaeologists found hundreds of people who had sought escape by sea. The 2001 study examined 80 of roughly 300 recovered people and proposed near-instant death from a surge around 500 °C. Later work reconstructed several surges at different temperatures. The remains document both the human toll and the eruption’s effects.

Rediscovery began underground. In 1738, Charles of Bourbon ordered shaft and tunnel excavations; work at the Villa of the Papyri began in 1750, and rolls like charcoal sticks were found in 1752. Counts vary by whether fragments are included; the Archaeological Park reports over 1,800 carbonized papyri. Greek philosophy predominates among opened works, especially texts by Epicurean Philodemus on ethics, poetry, rhetoric, music, vice and virtue. Latin works and other philosophers also appear, but the library’s full contents are unknown. UCLA’s Philodemus Project publishes editions and translations.

The villa was a grand seaside residence of the Roman elite, with sculptures as well as books, rather than a public library. Its owner has not been established; scholars have proposed a connection with the Piso family and with Philodemus, a first-century BC Epicurean philosopher. The prevalence of his works among opened scrolls tells us much about the known collection, but does not prove that every sealed roll is his or that the rest of the villa has been excavated. Much of it still lies beneath the modern town. The Herculaneum Society describes both the known collection and its uncertainties.

The discovery posed another historical problem: how to read objects that broke when touched. Early attempts included cutting scrolls lengthwise; in some cases one layer was copied and then scraped away to expose the next. Antonio Piaggio later devised a machine that slowly pulled sheets apart. These methods rescued some texts, but cost part of the material. The Herculaneum Society documents these methods. Multispectral photography subsequently improved reading of already exposed surfaces: wavelengths outside human vision can distinguish black ink from blackened papyrus. Even the best photograph of an outer surface cannot reach the sealed windings inside.

A book that cannot be opened

Papyrus is made of layers rolled together. Carbonization left it fragile, flattened, deformed and cracked. Pulling sheets apart can crumble them or tear away stuck layers; earlier attempts saved some text but also damaged scrolls. Some outer portions were already lost before modern study.

Instead of pulling at a sheet, researchers can pass X-rays through a scroll and reconstruct its interior. This approach requires two separate achievements: distinguishing tightly packed layers and detecting ink on a support that also contains carbon.

Light produced by electrons on a curve

The European Synchrotron Radiation Facility (ESRF) is in Grenoble, France. Early phase-contrast papyrus experiments took place there. Its BM18 beamline acquired the high-resolution scan behind the PHerc. 1667 result announced June 25, 2026, using a tuned protocol to separate layers and reveal faint signals.

Diamond Light Source in the UK contributed separate scans. Its I12 beamline examined PHerc. 172, held by Oxford’s Bodleian Libraries, in July 2024, helping recover text. A 2019 Diamond campaign scanned two scrolls and fragments from the Institut de France and recovered writing in one region.

An electric charge emits radiation when it accelerates. Changing direction counts, even at constant speed: a bicycle rounding a corner keeps its speed but changes direction. In a synchrotron, magnetic fields bend high-energy electrons circulating in a ring; undulators can make them oscillate. Their acceleration produces light, including X-rays, carried to experimental stations. Its brightness and controllability enable fine measurements difficult with conventional sources.

ESRF facility in Grenoble. Photo by Anders Sandberg, published on Wikimedia Commons; original page, CC BY 2.0 licence.

From many projections to a virtual sheet

Tomography resembles a medical CT scan: measurements from many angles are reconstructed into a 3D volume. Here, high-energy X-rays resolve micrometre details in small objects, rather than imaging a human body.

The volume shows fibres and layers, not a readable page. A papyrus surface must be located and segmented from the surrounding volume, with folds, breaks and stuck layers corrected. Geometry then virtually unfolds it into a 2D image while retaining each point’s position in the scroll.

Each point in the reconstructed volume is a voxel, the three-dimensional equivalent of a pixel. For PHerc. 1667, the 2026 study reports voxels about 2.4 micrometres wide in the principal scan. That number describes sampling, not a guarantee that two stuck sheets will always be distinguishable. Contrast, motion, reconstruction and compressed fibres also matter. The team tuned beam energy, sample-to-detector distance and phase-contrast processing to balance layer separation against ink visibility.

Geometry matters as much as resolution. A cross-section shows distorted nested lines, but following one of them along the whole scroll means avoiding jumps to the neighbouring winding. For PHerc. 1667, a 3D network proposed which voxels belonged to the writing surface. That prediction guided construction of a quadrilateral mesh; the team corrected doubtful regions by hand and checked the mesh against the original slices. They then spread the surface onto a plane with low distortion. Every pixel in the flattened image retains a corresponding 3D position: a doubtful letter can be checked in the original volume. To render the sheet, the team sampled CT values on either side of the surface along its local normal, rather than taking an arbitrary flat slice.

Original schematic of virtual unwrapping, simplified and not to scale. Point P on the mesh has coordinates in the volume (x, y, z) and on the flattened sheet (u, v); this mapping lets a reader return from a possible letter to the original measurement.

The contrast problem

Many ancient inks and carbonized papyrus are both rich in carbon, so absorption alone may not distinguish them. Some inks contain denser elements such as lead; recipes and preservation vary. There is no universal ink or imaging trick.

X-rays also shift in phase—like a wave being advanced or delayed—as they pass through materials. After travelling some distance, these shifts can appear as intensity differences, highlighting edges, fibres or subtle deposits. This helps locate layers and may reveal strokes, but does not make every ink visible; models sometimes detect weak patterns instead.

What the model learns

Different models serve two tasks: proposing the sheet surface and estimating where ink lies. Tomography is reconstructed from the projections; final flattening is a geometric operation on the reviewed mesh. None of these models translates Greek. To understand what ink detection adds, it helps to examine its basic operation: convolution.

What an image convolution does

A digital image is a grid of numbers. A small filter, or kernel, moves across it and computes a weighted sum of nearby values at each position. Imagine three pixels in a row with intensities 20, 20, 80. Applying weights −1, 0, +1 gives −20 + 0 + 80 = 60 at the centre: the filter responds to a change in intensity. A uniform region, 20, 20, 20, would give zero. This is a teaching example of edge detection, not a filter used by the project.

In two dimensions, the kernel is a small table, perhaps 3 × 3 values, that moves across the image’s height and width. In three dimensions, it is a small block that also moves between tomographic slices. A response can therefore depend on a fibre in one slice, its continuity in adjacent slices and a slight alteration on the surface. The weights in a convolutional network are learned during training, rather than written by hand as in the example, to reduce the difference between predictions and known ink or surface labels.

Invented example showing one position of a 2D filter: the nine pairs of values are multiplied and summed to give 140. Moving the filter would produce a complete feature map. In deep learning, this operation is commonly called a “convolution”, although without flipping the kernel it is mathematically a cross-correlation.

One layer of filters produces feature maps. Early layers may respond to local edges or textures; later layers combine those responses into broader patterns. Sharing the same weights across positions lets the network look for similar signals throughout a sheet. That advantage does not turn every mark into a letter: a crack, fibre or scan artefact can also activate filters. Carefully aligned examples, tests on unseen regions and review of the original volume are therefore essential.

From known fragments to hidden ink

In the initial EduceLab-Scrolls work, researchers aligned X-ray scans with infrared photographs, where ink is easier to distinguish. Specialists marked ink or background, then projected those labels onto the volume. A supervised model learned from small 3D patches of X-ray values. One published 3D convolutional network used context across slices to produce an ink-probability map.

Alignment is decisive: if a photographic mark falls a few pixels away from its CT location, the model may learn that a clean fibre is ink. The EduceLab-Scrolls study therefore excluded points close to uncertain label boundaries from training. It also tested predictions on fragment regions withheld from training and applied the detector to hidden layers of those fragments. No direct photograph exists of the hidden face being recovered.

Ink detection in a hidden layer of fragment F4, rendered from the X-ray volume. This is not a photograph of an intact scroll or an automatic reading of words. Figure from Stephen Parsons et al., EduceLab-Scrolls, CC BY-NC-SA 4.0; converted to WebP and shared under the same licence.

Explanatory diagram, not an exact representation of a specific architecture or scan. For real examples of strokes and their labels, see the documentation of the first ink discovery and this official training animation.

The more recent work on PHerc. 1667 used a different architecture: an encoder with 3D convolutions processes a small block aligned with the sheet, while a 2D decoder returns an ink probability at each surface position. The combination has a specific reason. Infrared photographs used as labels show where ink lies on the sheet, but not its exact depth in the volume. The model pools information across depth before producing the flat map. Its 256-pixel windows cover about 614 micrometres, smaller than a whole detected letter; this limits learning of letter or word shapes instead of local physical signals. The 2026 paper describes the architecture.

After training on fragments, the detector indicated strokes in sealed rolls absent from that initial training. The team selected promising regions, generated provisional labels from predictions and fine-tuned the model over several rounds. Such pseudo-labelling can make weak signals more visible, but can also reinforce errors if a prediction is accepted uncritically. For PHerc. 1667, columns were held aside to check progress, and the final images underwent papyrological review. Neither optical character recognition nor language models generated the inferred ink.

For evaluation, researchers compare predictions on unseen regions against labels: which strokes were recovered, and which marks were false? Early examples were scarce and some characters went undetected. Labels also depend on alignment, preservation and human judgement, so they are not perfect ground truth.

For PHerc. 1667, a 3D convolutional U-Net-style network proposes surface locations; its geometry becomes a mesh for inspection and correction. In PHerc. Paris 4, a separate 3D network segments ink deposits directly visible in the volume thanks to the improved scan. Their agreement with strokes recovered on the flattened sheet provides an independent physical check for that scroll; it does not mean ink is equally visible in every papyrus.

Different papyri, scanners and inks can change the signal. Folds, breaks, damage and faint strokes challenge tracing and detection. A model therefore produces a map of probable evidence, not a transcription. Papyrologists assess letters, uncertainty, Greek meaning and context; technical teams keep each proposal traceable to the data.

What has been read, and what remains

The June 25, 2026 Vesuvius Challenge announcement presented PHerc. 1667: about 1.4 metres of surviving written surface and 22 columns or equivalents, virtually unwrapped and reviewed by papyrologists. The preliminary paper, published two days later, notes that some traces remain unreadable and lost or uncertain letters follow papyrological conventions.

The text appears to be a philosophical treatise on ethics, human nature and impulse. The name Aristocreon, linked to Stoic Chrysippus, suggests a Stoic context and probable second-century BC date; author and title remain unknown. Separately, PHerc. 139 yielded evidence for Philodemus’ On the Gods, Book 8, and Diamond’s PHerc. 172 scan revealed On Vices, Book 1.

“Reading the whole scroll” means studying all surface that survives in PHerc. 1667, not recovering missing parts or an entire original book. Layers may be broken or carry ink without enough evidence to read it; each scroll has its own geometric and imaging limits.

Ruins at Herculaneum. Photo by Christofle Van Der Meulen; original page, dedicated to the public domain under CC0 1.0.

Explore the project

Official ink-detection tutorial: code and a reproducible walkthrough of training and inference for a project ink-detection solution.
Vesuvius Challenge data and formats: scans, volumes and technical documentation published by the project.
Official ScrollPrize/villa repository: the project’s monorepo, with tools for segmentation, virtual unwrapping and ink detection. It is not a single participant’s implementation.
2026 paper on PHerc. 1667: scientific preprint describing methods, limitations, papyrological review and materials for reproducing analyses.

The work makes most sense as a chain. Physics obtains a measurement without dismantling the object; computational reconstruction locates and flattens its sheets; machine learning helps highlight ink patterns in the data; and specialists turn those signals into historical readings with explicit levels of certainty. Publishing scans, labels, code and reviews lets other teams check every stage and suggest improvements.

Convolutional Networks on josiete.com