An X-ray can reveal that an artist painted over an earlier composition without cleanly separating the two images. A Northwestern-led team is developing AI to untangle those overlapping signals, using the physics of pigments and imaging to constrain its reconstructions. The Art Institute of Chicago and the Van Gogh Museum will contribute technical records, material expertise and paintings for the research.[1]
Northwestern announced the project in October as one of 19 initiatives sharing $10 million from the Schmidt Sciences Humanities and AI Virtual Institute. That is the programme-wide award total, rather than a grant of $10 million to this project alone.[1]
An X-ray contains an equation with too many unknowns
X-rays travel through layers and interact differently with different pigments and materials. In a finished radiograph, information from those layers overlaps. If two different sequences of paint can produce a similar measured signal, there may be no single image that the data force us to believe. This is an inverse problem: researchers see the combined result and try to infer the causes. Inverse problems occur in medical imaging and astrophysics too, but conservation adds complicated material mixtures and centuries of ageing.[1]
A generic image-generation model could produce a visually convincing underlying portrait. That may be the worst possible scientific outcome if the image acquires the authority of an archival discovery. The proposed Northwestern method instead incorporates physical knowledge about light, X-rays and pigments. A simulator will generate plausible layered samples; the AI will learn how different configurations could produce different measurements, and constraints from physics will reject reconstructions that cannot explain the recorded signals.[1]
The project plans to combine more than one imaging modality and quantify uncertainty. If a layer boundary is strongly supported by the evidence, a reconstruction can reflect that confidence. If two configurations remain equally plausible, the researchers want the system to display the ambiguity rather than selecting whichever image looks prettier. The challenge is to make the inference inspectable by conservators, who can compare it against pigment analysis, documentation and the known history of an artwork.[1]
The artists’ decisions, not just their secret pictures
A concealed drawing can reveal why a finished composition looks the way it does. It may show a figure moved, a colour changed, a pose abandoned or a canvas reused during a period of financial constraint. These decisions link artistic technique to biography, studio practice and the availability of materials. An AI reconstruction is useful when it clarifies a chain of evidence that historians can challenge, not merely when it generates a dramatic before-and-after animation.[1]
Van Gogh’s documented works offer one test case because the partner museums have extensive technical and historical records. A successful system would need to behave plausibly on synthetic examples and then survive validation against real paintings where multiple kinds of evidence can constrain the hidden layers. Public technical records and future scholarly scrutiny matter far more than a single promotional image.[1]
A reconstruction that preserves alternatives
The team plans to make its tools open source and present the work through a public exhibition. Where several hidden-layer arrangements fit the measurements, the framework is intended to show a range of possibilities and identify where the evidence is strongest.[1]
The aim is to recover decisions as well as images: a reused canvas, a moved figure or an abandoned composition. Keeping those reconstructions tied to measured materials would make them useful evidence for art history.[1]
Sources & further reading
- A New AI Framework to Reveal the Hidden Histories of Paintings Northwestern Engineering · 15 Oct 2026 · Accessed 24 Oct 2026