Feb 27 2018
A new technique feeds experimental measurements of a quantum system to an artificial neural network. The network learns over time and attempts to impersonate the quantum system’s behavior. With enough data, scientists can fully reconstruct the quantum system. (CREDIT: Giuseppe Carleo/Flatiron Institute.)
Methods similar to those used to train self-driving cars and chess-playing computers are at present assisting physicists in investigating the complexities of the quantum realm.
As a first ever attempt, physicists have showed that machine learning can be adopted to rebuild a quantum system by using comparatively lesser experimental evaluations. This technique will enable researchers to absolutely investigate particle systems exponentially faster when compared to traditional, brute-force methods. Complex systems that would need thousands of years to be rebuilt using earlier techniques could be thoroughly investigated within a few hours.
The researchers have described in Nature Physics on February 26, 2018, that the study will be advantageous for the advancement of quantum computers and other quantum mechanical applications.
We have shown that machine intelligence can capture the essence of a quantum system in a compact way. We can now effectively extend the capabilities of experiments.
Giuseppe Carleo, Associate Research Scientist, Center for Computational Quantum Physics, Flatiron Institute
AlphaGo inspired Carleo, who carried out the study while he was a lecturer at ETH Zurich. The computer program adopted machine learning to outperform the world champion of Go, a Chinese board game, in 2016.
AlphaGo was really impressive, so we started asking ourselves how we could use those ideas in quantum physics.
Giuseppe Carleo, Associate Research Scientist, Center for Computational Quantum Physics, Flatiron Institute
Particle systems such as electrons can occur in several distinctive configurations, each with a specific probability of occurrence. For example, each electron can have an upward or a downward spin, identical to Schrödinger’s cat being dead or alive in the well-known thought experiment. In the quantum world, unobserved systems do not occur as any one of these arrangements. In contrast, the system might be imagined to exist is all probable configurations at the same time.
Upon evaluation, the system crumbles into a single configuration, quite similar to Schrödinger’s cat, which is either alive or dead upon opening its box. This anomaly in quantum mechanics denotes that one can never thoroughly observe the complexity of a system in just one experiment. Rather, experimentalists carry out the same evaluations again and again until they can ascertain the state of the entire system.
That technique holds good for uncomplicated systems including just a few particles. However, “things get nasty with a lot of particles,” stated Carleo. With an increase in the number of particles, the complexity increases exponentially. By taking into account just the fact that each electron can have a spin up or down, a system including five electrons has 32 probable configurations, and a system with 100 electrons has over 1 million trillion trillion probable configurations.
The matters are further complicated by the entanglement of. Quantum entanglement results in intertwining of independent particles, which cannot further be considered to be purely separate entities though they are physically isolated. Such an entanglement modifies the possibility of various configurations.
Therefore, traditional techniques do not hold good for complex quantum systems.
Giacomo Torlai from the University of Waterloo and the Perimeter Institute in Canada, and also Carleo and his team overcame these drawbacks by adopting machine learning methods. The scientists fed experimental evaluations of a quantum system into a software tool based on artificial neural networks. Over time, the software learns and endeavors to imitate the behavior of the system. Upon ingesting adequate data, the software can precisely rebuild the entire quantum system.
The scientists investigated the software through simulated experimental datasets based on disparate sample quantum systems. In the investigations, the performance of the software far exceeded that of the traditional techniques. For a system with eight electrons that had spin up or down, the software was able to precisely rebuild the system with close to just 100 evaluations. In comparison, a traditional brute-force technique needed nearly one million evaluations to achieve the same precision level. The new method could also handle considerably larger systems. Conversely, this potential could assist researchers in validating whether a quantum computer is correctly constructed and whether any quantum software would perform as expected, suggested the scientists.
Reproducing the fundamental qualities of complex quantum systems with compact artificial neural networks has other well-extending outcomes. Andrew Millis, co-director of the Center for Computational Quantum Physics, indicates that the concepts offer a significant, innovative strategy to the center’s ongoing advancement of innovative techniques for gaining knowledge of the behavior of interacting quantum systems, and relate to studies on other quantum physics–inspired machine learning strategies.
Carleo stated that, apart from applications to basic research, the knowledge gained by the team by integrating machine learning with concepts from quantum physics could enhance general-purpose applications of artificial intelligence.
We could use the methods we developed here in other contexts. Someday we might have a self-driving car inspired by quantum mechanics, who knows.
Giuseppe Carleo, Associate Research Scientist, Center for Computational Quantum Physics, Flatiron Institute