German researchers develop AI to predict liquid properties
Artificial intelligence is reshaping how scientists investigate complex physical systems, with new advances enabling faster and more precise simulations of phenomena ranging from liquid behavior to electron dynamics in water.
Researchers at the University of Bayreuth announced this week that they have developed an AI based method capable of significantly accelerating the calculation of liquid properties. The findings, published in Physical Review Letters as an Editors’ Suggestion, describe an approach that predicts chemical potential, a key quantity for understanding liquids in thermodynamic equilibrium, without relying on the computationally intensive algorithms traditionally required.
The team, led by Professor Matthias Schmidt and Dr Florian Sammüller, introduced a neural network that does not directly learn the chemical potential itself. Instead, it learns what physicists call a universal density functional, a mathematical framework that captures the fundamental physical relationships within a liquid and remains valid across many different systems.
Schmidt said the distinguishing feature of the method is that the AI focuses on underlying physical principles rather than reproducing specific target values. Sammüller explained that the approach blends data driven learning with established theoretical physics. The researchers compared the concept to an image recognition system that could identify cats without ever having seen one during training. By embedding physical laws into the model, the system generalizes across different liquid systems with greater efficiency.
The scientists said the method could support computer aided design of new materials and pharmaceuticals by reducing the time and computing power required for simulations. Faster and more scalable liquid modeling may also improve research in chemistry, soft matter physics and energy storage.
Separate efforts are also pushing the boundaries of physics informed AI. A collaboration known as Polymathic AI has developed foundation models trained not on text or images, but on physical systems. The University of Cambridge announced in late January that two models, Walrus and AION 1, can transfer knowledge learned from one class of physical systems to entirely different problems.
Walrus, a transformer model with 1.3 billion parameters, was trained on 15 terabytes of data covering 19 scenarios and 63 physical fields, including astrophysics, geosciences, plasma physics and classical fluid dynamics. Dr Miles Cranmer from Cambridge’s Department of Applied Mathematics and Theoretical Physics said he was struck by the fact that a multidisciplinary physics foundation model can function across such diverse domains.
In the United States, researchers at Lawrence Berkeley National Laboratory reported another advance in January. Led by Alvarez fellow Pinchen Xie, the team developed a hybrid method combining quantum mechanics and machine learning to simulate electron behavior in water. The approach accurately predicts reaction rates and electron energies in interactions with hydronium ions while using far less computational power than conventional techniques.
Rafael Gómez Bombarelli, an associate professor at the Massachusetts Institute of Technology who has worked on AI driven materials discovery for more than a decade, described the field as reaching a second inflection point. He noted that scaling laws have proven effective in language models and in simulation tasks, and suggested that similar scaling strategies could now transform scientific research itself.
-
13:20
-
13:12
-
13:00
-
12:50
-
12:45
-
12:30
-
12:20
-
12:15
-
12:03
-
12:00
-
11:50
-
11:45
-
11:30
-
11:25
-
11:20
-
11:00
-
10:50
-
10:30
-
10:20
-
10:00
-
10:00
-
09:50
-
09:50
-
09:30
-
09:20
-
09:10
-
09:00
-
08:50
-
08:33
-
08:30
-
08:20
-
08:00
-
07:50
-
07:40
-
07:20
-
07:00
-
18:50
-
18:30
-
18:20
-
18:00
-
17:50
-
17:30
-
17:20
-
17:00
-
16:50
-
16:30
-
16:20
-
16:15
-
16:00
-
15:50
-
15:30
-
15:22
-
15:20
-
15:08
-
15:02
-
15:00
-
14:50
-
14:30
-
14:20
-
14:00
-
13:50
-
13:30
