Learnable program can independently plan and test new quantum experiments

Machine-brain with a spirit of research: Physicists have developed an artificial intelligence that can invent and test new quantum experiments. In this way, the computer does real scientific development work - instead of just evaluating the data from such experiments. Through virtual testing, the machine brain independently learns which approaches are sensible and which ones are not.

Learning-capable computer systems are almost ubiquitous: they are in our smartphone, launch tweets and articles on the net as webbots or help to expose fake news and fake images. Even in the judiciary, medical diagnosis and safety checks of nuclear power plants intelligent algorithms are now in use.

From the empty stage to the finished experiment

But there has been one exception so far: The invention and planning of scientific experiments and research work is still purely in human hands - still. For now, Alexey Melnikov from the University of Innsbruck in Austria and his colleagues have developed an artificial intelligence that can independently develop quantum experiments and try them out virtually.

At the beginning there is an empty laboratory table for photonic quantum experiments. The artificial agent now tries to develop new experiments by virtually mounting mirrors, prisms or beam splitters on the table. If his actions lead to a meaningful result, the agent notices this and relies on it for later attempts. The system stores many individual fragments of experience, which are networked together.

Learning from success and failure

The decisive factor here is that the intelligent system learns from success as well as failure and adapts its neural network accordingly.

When we analyse the results in the memory of the machine, we see that certain structures have developed, explains Hendrik Poulsen Nautrup from the University of Innsbruck.

In the first experiments, the system has already independently rediscovered experimental techniques which are standard in quantum optical laboratories.

At the same time, the machine brain can also create completely new scenarios and try them out.

The intelligent machine always looks for the best way to realize something and generates optimized experiments, explains Melnikov. And sometimes it also provides answers to questions we didn't ask.

This shows that machines could also play a creatively supportive role in research in the future.

Can AI be creative?

At present, artificial intelligence is still designed to solve a specific problem independently and experimentally. The researchers specify the task in each case. In the future, however, the scientists want to further expand the learning-capable program - it should then tackle entirely new tasks on its own.

But can a machine be more than just a tool? Will the AI of the future play a more creative role alongside the scientist? These are the questions that researchers also ask themselves. The role that artificial intelligence will play in laboratories in the future must still be shown - also through their work.

Source: Proceedings of the National Academy of Sciences, 23018; doi: 10.1073/pnas. 1714936115