USTC Releases "Machine Scientist 2.2": AI Moves From Running Experiments to Creating Substances

On October 18 in Hefei, at the fourth annual conference of the Intelligent Scientist Ecosystem Alliance, the University of Science and Technology of China (USTC) officially unveiled the Machine Scientist 2.2 series. Version 2.2 is positioned as the next step after lab automation: the system is supposed to create new substances on its own, not just execute predefined protocols. For now, this is the developers' claim. The available material contains no figures that would let anyone verify it, and that is worth keeping in mind while reading everything else.

The bar for expectations has moved

Jiang Jun, director of USTC's Center for Scientific Intelligence and Materials Creation, put the moment best. Five years ago, he says, many people who saw a robot running experiments thought: how clever. Today the reaction is different: "Is that all?" The reason, in his words, is large models, which have raised expectations of AI.

It is an accurate observation, and an uncomfortable one for the whole field. A robotic arm with a pipette stopped being news long ago. The market now expects something else from an "AI scientist": to propose hypotheses on its own, choose the next experiment, interpret the result and, ultimately, produce a substance that did not exist before. This is the direction in which USTC is pushing the Machine Scientist brand.

What is known about the release

Based on the available excerpt of the announcement, the picture is as follows:

• Product: Machine Scientist 2.2 series (机器科学家 2.2).
• Developer: University of Science and Technology of China, Center for Scientific Intelligence and Materials Creation.
• Announcement venue: the fourth annual conference of the Intelligent Scientist Ecosystem Alliance, Hefei, October 18.
• Stated goal: autonomous creation of new substances instead of automating individual experiments.
• Not disclosed in the available text: architecture, model composition, number of samples processed, number of confirmed new materials, timeline for access by outside labs.

The word "series" in the name is also telling: it apparently refers to a set of systems for different classes of tasks, not a single installation. That can only be confirmed through technical documentation or publications.

Context: from robot chemist to autonomous creator

The Hefei team has worked in this area for years. Jiang Jun's group previously showed a robot chemist that used published literature to select catalysts for producing oxygen from materials related to Martian meteorites. The emphasis then was on pairing AI with a robotic platform in a closed loop: the model proposes, the robot executes, and the data returns to the model.

Version 2.2 claims the next level. The difference between "optimizing the known" and "creating the new" is fundamental. In the first case, the system searches for an optimum within a defined space of compositions and conditions. In the second, it has to go beyond its training data without producing structures that cannot be made. This is where large models offer a chance: they handle literature better, formulate hypotheses and plan synthesis. They are also prone to confident errors, and in chemistry every error costs reagents and time.

The choice of venue is just as revealing. The Intelligent Scientist Ecosystem Alliance brings together participants who need shared standards, data and tools. If USTC is building an ecosystem around Machine Scientist rather than just a lab bench, the bet is on scaling through partners.

How to evaluate such claims

Until full data is available, it helps to keep a short checklist at hand. For researchers and investors, it looks like this:

• How many new materials the system produced on its own and that were confirmed by independent measurements.
• What share of decisions the model makes and what share a human makes: choosing the target, planning synthesis, analyzing failures.
• Reproducibility: whether a second run gives the same result on a different setup.
• Openness: whether data, an API or remote access to the platform will be available to outside groups.
• Applicability: batteries, catalysts, semiconductors or a narrow demonstration task.

If USTC publishes answers to at least the first two points, Machine Scientist 2.2 will become a notable reference point for labs and companies building their own autonomous platforms. If not, the release will remain a strong statement at a good conference. Watch for papers in peer-reviewed journals and for the alliance's list of partners: those will show whether the system works beyond a demonstration.