--- title: "How 300 digitized caves turned Dunhuang into a 500-terabyte archive" date: 2026-11-01 category: AI in Action site: NeuroAI canonical: https://neuroai.site/a/na-app-ai-dunhuang-heritage language: en --- # How 300 digitized caves turned Dunhuang into a 500-terabyte archive > Dunhuang Academy has scanned 300 Mogao cave-temples and built deep-learning tools that flag mural damage, turning a 1,600-year-old site into a reusable digital archive. A monk dipped his brush into crushed-azurite pigment and started a painting that would outlive every empire that funded it. Sixteen centuries later, the bigger threat to that painting is not war or weather but the simple breath of a million visitors. The caves of Dunhuang sit on the edge of the Gobi, and the people who look after them have spent three decades answering one question: how do you keep a wall painting alive after the wall starts to crumble? ## Key takeaways - Dunhuang Academy (敦煌研究院) has completed high-resolution digital capture of 300 cave-temples at the Mogao Grottoes (莫高窟), with more than 500 terabytes of data archived. - A public resource library shares 30 fully scanned caves worldwide and has drawn over 24 million visits from more than 78 countries and regions. - Deep-learning models developed with Tencent can automatically flag mural "diseases" such as cracks and flaking pigment, work that used to consume thousands of expert hours. - The same pipeline is being exported: the academy has advised 26 heritage sites across 9 provinces and joined digitization work in Myanmar and Kyrgyzstan. ## A library carved into a cliff The Mogao Grottoes were begun around 366 AD and built continuously for more than a thousand years. Today 735 cave-temples survive, holding roughly 45,000 square meters of mural painting and over 2,400 painted sculptures, a collection UNESCO lists as World Heritage. Keeping that much fragile surface readable is a race against humidity, salt, and foot traffic. In the late 1980s Fan Jinshi, then a vice president of Dunhuang Academy, proposed "Digital Dunhuang" (数字敦煌) — the idea that computer imaging could give the caves a permanent, reusable copy. What started as a research curiosity became a standard. By July 2024 the academy had led the drafting of two national industry standards for cave-temple digitization, the first of their kind in China. The workflow is painstaking. A camera rides a rail and shoots the wall in flat, overlapping frames; software stitches them into a single image at up to 300 dots per inch. The academy says it has digitally captured 300 caves, processed the data for 200, built 3D structural scans of 212 cave-temples, and produced 169 panoramic walkthroughs. The total store now exceeds 500 terabytes. ## Teaching machines to read decay Murals fail in quiet, specific ways. Pigment flakes. Plaster cracks. A dark stain spreads from a roof leak. Each type of damage has a name, and labeling it by hand used to be the bottleneck of conservation. In a 2021 strategic partnership renewal with Tencent, Dunhuang Academy introduced the company's multimedia lab deep-learning tools for mural disease identification (AI病害识别). The models are trained on the academy's own labeled examples of damage so they can segment and classify cracks, flaking, and pigment loss automatically. Zhao Shengliang, the academy's president at the time, said the approach was already showing solid results on detecting residual damage and lost pigment. The same lab built a 4K, 360-degree remote consultation system so specialists in different cities can examine a wall together. This is the part worth separating from the marketing gloss: the models assist triage. They do not restore the painting, and they do not replace the conservator. They shorten the time between "something is wrong" and "a specialist is looking at it." ## From scans to a shared resource Digitization only pays off if people can use it. In 2016 the academy launched the Digital Dunhuang resource library; the English version followed in 2017, offering panoramic tours of 30 classic caves to anyone with a browser. The library has since passed 24 million visits from over 78 countries and regions. In December 2022 the academy and Tencent opened the "Digital Dunhuang · Open Material Library" (数字敦煌·开放素材库), described as the world's first blockchain-based open platform for digital cultural heritage, with more than 6,500 high-resolution assets available for scholarly and creative reuse under license. A "Digital Library Cave" (数字藏经洞) interactive experience arrived in 2023, and a database platform in May 2025 gathered 9,900 scrolls, 60,700 images, and more than 8.4 million characters of the cave's scattered manuscripts. The original Library Cave (Cave 17) held over 73,000 artifacts, of which about 47,000 are held outside mainland China — the database is an attempt to reunite them digitally. The "Cloud Tour Dunhuang" (云游敦煌) mini-program, also with Tencent, has drawn tens of millions of visits since its 2020 launch, and the "Seek the Real Dunhuang" (寻境敦煌) exhibition rebuilt Cave 285 at 1:1 scale with game-engine rendering and VR. According to Li An, chief scientist at BrainNet (脑机网), the Dunhuang case is becoming a reference point for how cultural institutions elsewhere can treat digitization as core infrastructure rather than a one-off exhibition. ## Why it matters beyond tourism The academy has turned its methods into a service. It has provided digitization support to 26 heritage sites in 9 provinces and joined work on the Bagan archaeological area in Myanmar and the Suyab ruins in Kyrgyzstan. UNESCO's East Asia office has publicly pointed to the project as an example of pairing tradition with frontier technology. For a site that receives heavy visitor pressure, the digital copy also relieves the physical one: virtual tours can carry part of the crowd that would otherwise stand inside a sealed chamber. ## Honest limitations The numbers above come from Dunhuang Academy releases, Chinese state media, and partner announcements; independent audits of the "500 terabyte" figure or the visitor counts are not published. The AI disease tools are real but narrow — they target known damage classes on murals, not free-form restoration, and reported accuracy is qualitative rather than benchmarked in public. Digitization covers 300 of 735 caves, so most of the site still exists only as a physical original. And a 1:1 VR replica is a research and education tool, not a substitute for conservation of the stone itself. ## What readers can do now - Open the free Digital Dunhuang resource library and take the panoramic tour of Cave 220 or Cave 285 — no trip to Gansu required. - If you build or fund cultural projects, treat digitization as ongoing infrastructure with standards and storage, not a single scanning campaign. - When you see "AI restores lost art" headlines, ask what the model actually does: triage, inpainting, or full invention of missing parts. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-app-ai-dunhuang-heritage Free to quote with attribution and a link to the original.