Learning chemistry begins with recognizing patterns. With the guidance of a patient professor or a vast training dataset, both humans and AI can learn which atoms tend to bond together, how molecular structures determine a material’s properties and the conditions under which chemical reactions occur.
AI has one obvious advantage: scale. It can absorb and analyze more information than any human could in a lifetime. Scientific discovery, however, demands something more. It requires understanding *why *those patterns exist and applying that understanding to unfamiliar systems.
Can AI learn the scientific concepts that enable new discoveries? That question motivates a new study by two longtime collaborators at USC Viterbi’s Mork Family Department of Chemical Engineering & Materials Science and the Thomas Lord Department of Computer Science, within USC Viterbi and the USC Stevens School of Computing and Artificial Intelligence.
Ken-ichi Nomura, associate professor of chemical engineering and materials science practice, and Aiichiro Nakano, professor of computer science, physics and astronomy, and biological sciences, recently published their research findings in Nature Communications. The paper was co-authored with Priya Darshan Vashishta, Fluor Chair in Engineering and professor in chemical engineering and materials science, biomedical engineering, computer science, and physics and astronomy, and Rajiv Kalia, professor of physics and astronomy, computer science, chemical engineering and materials science, and biomedical engineering.
“The paper demonstrates that a powerful AI model is capable of independently learning one of chemistry’s foundational concepts, the chemical bond, even though it had never been explicitly taught what a chemical bond is,” Nomura explained.
The study builds on the researchers’ previous breakthrough, Allegro-FM, a foundation AI model capable of simulating interactions among billions of atoms across 89 elements of the periodic table. Allegro-FM demonstrated what a sufficiently large AI model could predict. This year’s study asks a different question: what had the model learned?
Opening the black box #
Chemical bonds are the glue that holds matter together. They determine how molecules form, why materials behave as they do, and how chemical reactions unfold. For more than a century, chemists have relied on the concept to explain phenomena ranging from medicines and batteries to stronger concrete and cleaner energy technologies.
None of that knowledge was ever explicitly taught to Allegro-FM. Instead, the researchers trained the AI model to predict the quantum-mechanical energies and forces governing interactions between atoms. The outputs were remarkably accurate – but had the AI simply become exceptionally good at recognizing statistical relationships, or had it learned one of the concepts that scientists use to understand the physical world?
Knowing whether the AI had learned a scientific concept required solving another problem. Modern AI models are often described as black boxes. Researchers can evaluate the answers they produce, but they usually cannot see what the models have learned internally or how they arrived at those answers. “For us, the challenge was: how do we show that AI has learned a concept, when AI is a black box?” Nakano said.
The researchers’ answer was a new analytical framework called Edge-wise Emergent Energy Decomposition (E3D). Nakano describes it as a form of computational imaging. Rather than treating the neural network as an opaque system, E3D allows researchers to examine how information moves through the model as it learns. By tracking how the AI distributes energy between neighboring atoms, the framework reconstructs the chemical information encoded inside the model itself.
E3D indicated that the AI had independently acquired a transferable concept of chemical bonding. It had learned one of chemistry’s most fundamental organizing principles – not by memorizing bond energies or bond types, but by inferring the concept from the underlying quantum-mechanical data.
To test that conclusion, the researchers used the model’s internal calculations to estimate bond-dissociation energies – the energy required to break a chemical bond. Those estimates closely matched experimentally established values, even though the model had never been trained using bond-energy data.
Beyond better predictions #
A prediction tells researchers what is likely to happen. A scientific concept helps explain why it happens and can often be applied far beyond the examples from which it was learned. “We know AI can answer what is known,” Nakano said. “But the frontier is really: can AI help discover something new? To do that, we need to understand whether AI is understanding concepts.”
That question has become increasingly important as AI systems continue to grow in size. Researchers have long known that larger models trained on larger datasets become dramatically more capable, even when no new algorithms are introduced. Yet the reason why is still unclear.
This study points toward one possible explanation. As models scale, they may begin to develop abstract scientific concepts rather than simply storing increasingly large collections of examples. AI researchers describe these unexpected capabilities as emergent abilities – behaviors that appear only once models become sufficiently large.
“First we need to understand how AI represents a scientific concept,” said Nomura. “Once we understand how those concepts connect to one another, we can begin exploring entirely new discoveries. That’s the future direction.”
Published on August 18th, 2026
Last updated on August 18th, 2026