AI Model Adheres to Physical Laws in Chemistry

MIT's AI model FlowER predicts chemical reactions accurately, conserving mass.

Researchers at MIT have developed an artificial intelligence model called FlowER. It predicts the outcomes of chemical reactions with high accuracy, adhering to the fundamental law of mass conservation.

According to the study’s authors, previous attempts to use LLM for these tasks had limited success. AI often ignored physical laws and could “create” new atoms or “remove” existing ones during calculations.

“It’s akin to alchemy,” explained researcher Junyoung Jung.

He noted that unlike ChatGPT, their model not only considers a set of inputs and outputs but also tracks all stages of substance transformation.

To address the issue, the team employed a method proposed in the 1970s—a matrix of bonds and electrons. This approach allows the program to track every charged particle in a reaction, ensuring no atom is created or lost.

The FlowER model was trained on a database from the US Patent Office, which contains over a million chemical reactions. According to the developers, the system already demonstrates equal or higher accuracy compared to existing counterparts.

Professor Connor Coley noted that the result of the work currently only proves the concept. The model does not yet cover some reactions involving metals and catalysts, but the team is already working on this.

“We are incredibly excited to achieve such reliable predictions of chemical mechanisms. The model conserves mass and electrons,” added Coley.

The project is entirely open. The code, models, and datasets are published on GitHub. The developers believe that FlowER is already a useful tool for evaluating reactions.

In the future, the model may find applications in pharmaceuticals, the search for new materials, and electrochemical systems.

Back in 2026, the Chinese technology company Kaiwa Technology from Guangzhou plans to unveil a humanoid robot with an artificial womb in the abdominal area.