Scientists from the ETH Zurich Department of Physics, together with researchers from the Hebrew University of Jerusalem, have designed a novel machine-learning algorithm. The platform analyses large data sets which describe a physical system. The machine then learns to extract these patterns from them the essential information required to interpret the underlying physics. Where the “intelligence” part comes in is with the algorithm being able to extract the relevant physical entities without having any prior knowledge of the connectivity pattern.
Recently machine learning has begun to be applied to physics problems. This is often for the classification of physical phases and the numerical simulation of ground states. Perhaps the most well-known example is CERN, which handles so much data in a single run of the Large Hadron Collider that it is physically impossible for human beings to manually check through the data for anomalies.
In the new development a machine has been designed which is not simply a numerical simulator or a ‘hypothesis tester’, but is instead an integral part of the physical reasoning process. The new algorithm uses data about a physical system and extracts important degrees of freedom, between atoms, that are most relevant to describe the system and which will provide the most relevant clues for understanding physical behavior.
The artificial intelligence has been published in the journal Nature Physics. The research paper is titled “Mutual information, neural networks and the renormalization group.”
In related news, chemists have developed an algorithm, produced through machine learning, that has succeeded in analyzing every known organic chemical reaction. The aim is to use the artificial intelligence for drug discovery. See: “AI has analyzed every chemical reaction ever performed.”
