Researchers at the University of Bayreuth have developed a novel artificial intelligence method that dramatically accelerates the calculation of liquid properties by learning the fundamental physical relationships governing liquids. The breakthrough centers on predicting the chemical potential—a key thermodynamic quantity essential for describing liquids in equilibrium—using a machine learning approach that integrates deep theoretical insights from physics. Traditional AI methods, such as supervised learning, rely on training models to directly predict specific outcomes by analyzing large datasets. For example, in image recognition, a neural network learns to identify cats by being shown thousands of labeled images. However, this direct approach faces major challenges when applied to complex physical quantities like the chemical potential. Computing it typically requires resource-intensive simulations, making it impractical for rapid or large-scale analysis. To overcome this, Prof. Dr. Matthias Schmidt and Dr. Florian Sammüller from the University of Bayreuth designed a new AI framework that bypasses direct prediction. Instead of learning the chemical potential itself, the neural network learns the universal density functional—a mathematical description that captures the intrinsic physical behavior of liquids and soft matter systems. This functional encodes the fundamental laws that govern how particles in a liquid distribute themselves in response to external forces, regardless of the specific system. What makes the method unique is that it leverages the universality of these physical laws. For instance, when the same liquid is placed on different patterned surfaces, the underlying physics remains consistent. The AI learns this universal behavior from data across various scenarios, enabling it to generalize effectively to new systems without retraining. Once the density functional is learned, the model uses it to compute the system’s observable properties—such as particle density profiles and external potentials. The difference between the predicted and actual values is not filled in by the AI, but resolved through established principles of thermodynamic stability. From these principles, the chemical potential can be derived uniquely and consistently. “This approach combines data-driven machine learning with deep physical insight,” says Sammüller. “The AI learns the universal rules of liquid behavior, and the final result—chemical potential—is obtained through physical consistency, not direct training.” In essence, the AI learns the rules of the game, not the outcome. The researchers describe their method in a new study published in Physical Review Letters. They highlight that their technique is not only faster than conventional simulation methods but also more reliable and generalizable. It represents a significant step toward integrating AI with fundamental physics to solve complex problems in materials science, chemistry, and fluid dynamics. In analogy to image recognition, the method is akin to teaching an AI to identify cats without ever showing it a single cat—by teaching it the universal principles of what makes something a cat, rather than memorizing examples.