Researchers at USC Viterbi have found evidence that a powerful AI model independently learned the concept of a chemical bond - one of chemistry's most fundamental ideas - without ever being explicitly taught what a bond is. The findings, published in Nature Communications, offer a rare look inside the "black box" of a large AI model and suggest that scaling AI may lead to genuine conceptual understanding, not just pattern recognition.
The study builds on Allegro-FM, a foundation AI model developed by the same team that can simulate interactions among billions of atoms across 89 elements. Earlier work showed what the model could predict. This study asks what the model had actually learned.
Looking inside the model
Chemical bonds determine how molecules form, why materials behave as they do, and how reactions unfold. Chemists have relied on that concept for more than a century to explain everything from medicines to battery technology. But none of that knowledge was explicitly programmed into Allegro-FM. The researchers trained it to predict quantum-mechanical energies and forces between atoms. Its outputs were accurate. The question was whether the model had simply gotten good at spotting statistical patterns or had grasped a scientific concept.
That question creates a puzzle. Modern AI models are often described as black boxes: researchers can see the answers they produce but not what the models know internally or how they arrive at those answers.
"For us, the challenge was: how do we show that AI has learned a concept, when AI is a black box?" said Aiichiro Nakano, professor of computer science, physics and astronomy at USC.
The team's answer was a new analytical framework called Edge-wise Emergent Energy Decomposition, or E3D. Nakano describes it as a form of computational imaging. Instead of treating the neural network as an opaque system, E3D tracks how information moves through the model as it learns. By observing how the AI distributes energy between neighboring atoms, the framework reconstructs the chemical information encoded inside the model.
What the model learned
E3D indicated the AI had independently developed a transferable concept of chemical bonding. It had not memorized bond energies or bond types. It inferred the concept from the quantum-mechanical data itself.
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, despite the model never having been trained on bond-energy data.
"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," said Ken-ichi Nomura, associate professor of chemical engineering and materials science practice at USC.
What large models may be doing
The study points toward a possible explanation for why larger AI models trained on larger datasets become dramatically more capable even when no new algorithms are introduced. As models scale, they may begin to develop abstract scientific concepts rather than simply storing more examples. AI researchers call these unexpected capabilities emergent abilities.
"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."
The result doesn't just improve prediction accuracy. A prediction tells a researcher what is likely to happen. A scientific concept helps explain why it happens - and can be applied to systems the model has never encountered.
Why this matters for scientists and researchers
For researchers using AI as a tool, this study suggests that models trained on physical data may develop useful conceptual knowledge even when they are not explicitly taught it. That opens the possibility of using large AI models not merely as fast predictors to make fully trained to make discoveries: if a model has learned what a chemical bond is, researchers may be able to probe how it organizes other scientific concepts - and potentially surface insights that humans have not yet articulated.
"First we need to understand how AI represents a scientific concept," Nomura said. "Once we understand how those concepts connect to one another, we can begin exploring entirely new discoveries. That's the future direction."
Your membership also unlocks:
