The development of a new theory is often associated with great physicists. For example, you can think of Isaac Newton or Albert Einstein. Many Nobel Prizes have been awarded to new theories. Researchers at Forschungszentrum Jülich have now programmed an artificial intelligence that can deal with this too.

Their AI can recognize patterns in complex data sets and formulate them in physical theory. The development of a new theory is often associated with great physicists. For example, you can think of Isaac Newton or Albert Einstein. Many Nobel Prizes have been awarded to new theories. Researchers at Forschungszentrum Jülich have now programmed an artificial intelligence that can deal with this too. Their AI can recognize patterns in complex data sets and formulate them in physical theory.

In the following interview, Professor Moritz Helias from the Forschungszentrum Jülich Institute for Advanced Modeling (IAS-6) explains what “artificial intelligence physics” is and how it differs from traditional approaches.

How do physicists arrive at a new theory?

You usually start by observing the system before trying to propose how various components of the system interact with each other to explain the observed behavior. New predictions are then derived and tested based on this. A famous example is Isaac Newton’s law of gravity. Not only does it describe the gravitational force on Earth, but it can also be used to fairly accurately predict the motions of planets, moons, and comets, as well as the orbits of modern satellites.

However, the ways to arrive at such hypotheses are always different. You can start with the general principles and fundamental equations of physics and derive a hypothesis from these, or you can choose a phenomenological approach and limit yourself to describing observations as precisely as possible without explaining their causes. The difficulty lies in choosing a good approach among many possible approaches, adapting and simplifying if necessary.

What approach do you take regarding artificial intelligence?

Broadly speaking, this involves an approach known as “physics for machine learning.” In our working group, we use physics techniques to analyze and understand the complex function of artificial intelligence.

The important new idea, developed by Claudia Merger from our research group, was first to use a neural network that learns to accurately represent observed complex behavior in a simpler system. In other words, artificial intelligence aims to simplify all the complex interactions we observe between system components. We then take a simplified system and create a reverse mapping using trained AI. We develop a new theory by turning from a simplified system to a complex system. On the return path, complex interactions are partially composed of simplified interactions. In the end, the approach is not much different from that of a physicist, except that the way interactions are put together is now read from the parameters of the artificial intelligence. This view of the world, describing it in terms of interactions between its various parts that obey certain laws, is the basis of physics, hence the term AI physics.

In which programs was artificial intelligence used?

For example, we used a dataset of black-and-white images with hand-written numbers, which are often used in neural network research. As part of her doctoral thesis, Claudia Merger investigated how small substructures, such as the edges of numbers in images, are formed from interactions between pixels. Groups of pixels are detected, which tend to be brighter together and thus affect the shape of the edge of the number.

How big is the computational effort?

Using artificial intelligence is the trick that makes calculations possible in the first place. You get a huge number of possible interactions very quickly. Without using this trick you may only be looking at very small systems. However, even in multicomponent systems, the computational effort is still high due to the large number of possible interactions. However, we can efficiently parameterize these interactions so that we can now display systems with approximately 1000 interacting components, i.e. image regions of up to 1000 pixels. Much larger systems will also be possible with further optimization in the future.

How does this approach differ from other AIs like ChatGPT?

Many AIs try to learn the data theory used to train the AI. But the theories that AI learns are often uninterpretable. Instead, they are implicit in the parameters of the trained AI. On the contrary, our approach isolates the theory under consideration and formulates it in the language of interaction between the components of the system, which is the basis of physics. It therefore belongs to the field of explainable AI, specifically the field of “AI physics” because we use the language of physics to explain what AI learns. We can use the language of interaction to build a bridge between the complex inner workings of artificial intelligence and the theories that humans understand.