24/09/2026
Every few weeks, another model arrives with better benchmarks and new capabilities, and it’s easy to assume everything else should now move at the same pace.
Science and engineering are benefiting from that speed too. In materials research, AI can now suggest new materials at a scale no human team could manually explore: one recent effort produced 2.2 million candidate crystals, including around 380,000 predicted to be stable. Newer systems are also starting to build basic chemical rules into the process, so fewer proposed materials turn out to be unusable.
The next challenge is finding which materials are actually useful for a specific technology. In a study published in ACS Nano, XPANCEO researchers worked with Nobel laureate Konstantin Novoselov, Chair of the company’s Scientific Advisory Board, to screen tens of thousands of computer-predicted atomically thin crystals for materials that bend light most strongly, a key property for guiding light compactly on a chip.
The findings point researchers toward the material families worth testing in the lab. That experimental loop is also starting to change: at Rice University, robots will run experiments while an AI assistant learns from every result and helps choose what to test next.
The pattern becomes obvious quickly: the digital part can move much faster than the physical one.
In software, an idea can become code quickly, then be copied, rewritten, tested, and released as a new version almost immediately. Hardware still has to go through the real world every time: a material has to be synthesized, a component has to be fabricated, a device has to be assembled, and everything has to be tested under real conditions.
So it doesn’t make much sense to judge hardware by the same clock as software. Physical experiments take longer, but they also leave something behind: data, manufacturing knowledge, better processes, and a much clearer idea of what is worth trying next.
AI can help researchers decide which experiments are worth running and rule out weak directions before they reach the lab. It doesn’t remove the slower physical process, but it can make each step much more useful.
Software has trained us to expect almost instant iteration, but deep tech is slower for a simple reason: at some point, you have to make the thing.
