A large, controlled study at an unnamed US company suggests that AI can meaningfully change the pace of materials discovery. According to research conducted by MIT economist Aidan Toner-Rodgers, teams using a customized AI research tool discovered 44% more new materials than teams that continued with standard workflows.
The result is not just about faster idea generation. The AI-supported teams also filed 39% more patent applications, pointing to a measurable effect on work that can move from the lab toward protected inventions.
What the study examined
The study looked at a company with over 1,000 researchers. The company develops inorganic materials, including molecular compounds, crystal structures, glasses, and metal alloys.
Those materials are used for healthcare, optics, and industrial manufacturing. That matters because materials research is not a purely abstract exercise. Researchers are looking for structures with properties that can eventually be tested, refined, and used in product prototypes.
In the study, some research teams were randomly selected to use the customized AI tool. Other teams kept working with standard workflows. That setup made it possible to compare AI-supported research with more conventional research practices inside the same broad organization.
How the AI research tool worked
The system combined graph neural networks with reinforcement learning. It was pre-trained on data from extensive databases, including the Materials Project for crystal structures and the Alexandria Materials Database for molecular structures.
The workflow began with researchers entering desired material properties into the neural network. The tool then suggested new structures that might have those properties.
From there, the work still depended on human research teams. Scientists had to decide which suggestions looked worth pursuing, filter out likely failures, attempt to synthesize promising structures, and test them in experiments. Some candidates were also tested in product prototypes.
The process formed a feedback loop. Results from experiments and prototype work were fed back into the neural network, improving its predictive capabilities over time.
Why expertise still mattered
One of the most important findings was that the benefits were not evenly distributed. The company’s highest-performing researchers gained the most from AI assistance. Lower-performing scientists saw little advantage.
Toner-Rodgers suggests a practical reason for that gap. Top researchers may be better at judging which AI-generated suggestions are worth advancing. Their expertise helps them prioritize promising candidates and avoid wasting effort on weak leads.
By contrast, others may spend more resources testing false positives. In that case, the AI tool still produces suggestions, but the researcher’s ability to separate useful options from poor ones becomes the bottleneck.
This makes the study more nuanced than a simple story about automation replacing scientific judgment. The AI system produced candidate structures, but researchers still had to interpret, select, synthesize, test, and learn from the results.
Productivity came with a tradeoff
The productivity gains were clear in the reported outcomes: 44% more new materials and 39% more patent applications for teams using the AI tool. But the follow-up survey showed a less comfortable side of the shift.
Researchers using the AI reported lower job satisfaction. The reason given in the source was that the tool took over some of the more creative steps in their work.
Instead of spending as much time generating possible materials themselves, scientists mostly had to choose which suggested materials should move to the next phase. That may increase output, but it can also change how the work feels to the people doing it.
What this means for materials discovery
The study points to a future in which AI-supported research can expand the number of viable candidates moving through a materials pipeline. In fields involving inorganic materials, that could mean more structures are proposed, tested, and potentially advanced toward practical applications.
At the same time, the findings show that AI is not a universal productivity boost by itself. The strongest results came when capable researchers used the tool well. The system’s value depended not only on its predictions, but also on human judgment about which paths deserved time and resources.
For organizations working in materials discovery, the lesson is likely to be operational as much as technical. An AI research tool can increase discovery and patent activity, but it also changes the role of the researcher. The people using it may spend less time inventing from scratch and more time evaluating machine-generated options.
That shift can be powerful, but it is not neutral. The same tool that helps teams find more new materials can also make scientific work feel less creative. The study’s central message is therefore balanced: AI can raise research productivity, but the best results still depend on expert selection, experimental follow-through, and attention to how the work changes for researchers themselves.