Artificial Intelligence (AI) is one of the fastest growing areas of science and technology. It is not only a field in its own right helping to expand the capabilities of computer-based learning, but is also a powerful tool that, when applied to science and technology research, has the potential to transform how scientists conduct experiments.

In astronomy, for example, where the sheer volume of data generated calls for a new paradigmatic approach, AI is used to search for galaxy clusters in telescope images containing millions upon millions of stars. At the opposite end of the scale, it is also exploited in biochemical and biophysical research to understand complex microscopic processes.

Scientists at large international experimental facilities are also keen to reap the benefits of using AI and machine learning to enhance their research. The quantity of data collected on neutron and x-ray instruments is currently increasing exponentially, while only a fraction of this data is actually analysed properly. AI has the potential, on the one hand, to improve measurement strategies and, on the other hand, to help scientists identify keys features in their data rapidly.

The data challenge

While the capability and scope of AI research continues to grow, so too do the challenges facing large-scale facilities in producing, handling, treating and fully exploiting their data. Technological advancements often lead to the generation of ever greater volumes of increasingly complex data, making the task of extracting scientific insights all the more difficult.

Research based on neutron and x-ray scattering at major facilities is often conducted by visiting scientists, who travel from all over the globe to make use of world-class instruments and carry out their experiments. These external scientists receive invaluable assistance from the in-house neutron or x-ray experts responsible for the instruments in order to set up and conduct their experiments and interpret the data they obtain. However, these experiments generate vast amounts of raw data, much of which may go unanalysed.

Joining forces to make use of Artificial Intelligence

At the joint invitation of the ILL, the European Synchrotron Radiation Facility (ESRF) and the UK’s Science and Technology Facilities Council (STFC), experts from around the world gathered at the EPN Science campus in Grenoble, France, to discuss the potential of AI and machine learning in tackling the unique challenges related to photon and neutron science.

The workshop triggered global interest, reaching its maximum capacity with over 150 attendants from the major institutes using neutrons, x-rays and muons, including the Diamond Light Source, ISIS neutron and muon source, Oak Ridge National Laboratory (ORNL), and the Shanghai Synchrotron Radiation Facility, as well as of course ESRF and ILL. On top of this, nearly 2,600 individuals from around the world connected to the live stream. Lively discussions took place with insights on what is currently going on in exploring AI potential applications.

A unique context

While the benefits of using AI and machine learning in any scientific field may seem obvious, facilities such as the ILL present a number of unique challenges. In particular:

• The amount of usable data is often not sufficient. Usable data is data with necessary metadata so the AI can make connections and form conclusions. It also covers past data with known conclusions – as this data can then help us to train an AI, so that it can recognise these datasets for what they are. For example, knowing what the structure of the sample and optimal setup is for images helps us to train an AI to recognise features within them.
• The production of data directly at the instruments is extremely expensive and time-consuming, limiting the resources available for the development of AI-based approaches.
• Metadata containing explicit information on all possible experimental conditions (for example temperature, pressure, sample composition and orientation) are often missing or incomplete, limiting the possibility to use the data set for neural network training.
• For the above reasons, most AI training takes place using simulated data, making it essential to have reliable models of the physical systems.

At the ILL, we are already exploring techniques that will help maximise the neutron’s potential to enhance scientific understanding. ILL was one of the first scientific user facilities to implement DOI for its scientific users. It is leading in the setting of new standards for the handling and accessibility of scientific data. We are applying advances in digital technologies such as AI to improve the handling and usability of scientific results, and developing world-leading software solutions – in particular for the management of Digital Object Identifiers for data – to guarantee the traceability of scientific results from production through to publication.

The experimental technique we chose for our initial exploration of how to use AI for neutron scattering is Small-Angle Neutron Scattering (SANS), which can be used to deeply probe both soft and hard matter covering everything from biological molecules to crystal materials. This technique is deployed on a number of instruments at the ILL, including D22, which, among other things, has been used for experiments aimed at enhancing our understanding of how the molecules involved in type-2 diabetes develop, or exploring the potential of silkworm proteins for drug-releasing wound dressings.

To speed up the rate at which experiments are conducted and to democratise access by enabling scientists to carry out research more easily, we have begun to create a prototype of a neural network capable of identifying the sample structure and predicting the optimum setup for the measurements.

The system has already demonstrated good prediction capabilities and now the network knowledge must be extended to more complex structures. By further developing this programme and other similar techniques, we could in the near future relieve our users of some of the most repetitive and tedious tasks.

The future of AI for scientific discovery

AI will undoubtedly enable faster data reduction, in real time, and the integration of experiments, simulations and data analysis into a single seamless scientific project. This could in the long term lead to highly automated measurements or to the development of programmes that link past, present and future experiments by scientists spread across the globe, allowing them to extract meaningful insights from previously unconnected datasets.

For the time being, however, more collaboration, the sharing of datasets and coordinated activities to apply AI to neutrons and x-rays will be the key to bringing the most powerful computational techniques to the world’s most powerful analytical tools.

Food for thoughts

• What aspect of working with international experimental facilities might be benefit the most from artificial intelligence?
• How should scientific data management practices be changing in order to better support scientific and technological progress?

By Dr Paolo Mutti, Head of Scientific Computing at the Institut Laue-Langevin (ILL)