AI-Powered Autonomous Lab Creates 41 New Compounds in Just 17 Days

⇧ (VIDEO) You may also like this partner content

A-Lab, an AI-powered autonomous laboratory, synthesized 41 new powdered solid inorganic compounds in just 17 days, based on data from the Berkeley Lab Materials Project (the world’s largest database of inorganic compounds) and other AI. specializing in material design powered by Google DeepMind. The platform accelerates the discovery of new materials more than ever and promises broad application prospects.

Although the creation of new materials can now be done on a large scale using high-performance computers, experimental testing is difficult and time-consuming. Reducing this experimentation time requires automation and a certain amount of autonomy. Although research efforts in this direction are increasingly moving towards machine learning, a gap remains when it comes to training. Researchers at Berkeley Lab (at the University of California, Berkeley) suggest that this autonomy will require synergy between encoded data, data from different already established sources, and active AI learning.

The A-Lab laboratory was designed with this vision, combining robotics with databases from the beginning, data interpretation through machine learning, synthesis of available scientific literature and active learning. Specifically, the AI ​​that powers the robotic system creates “new recipes” by scanning scientific publications and makes adjustments using active learning. Materials Design Data and
(the specialized AI developed by Google DeepMind) made it possible to predict the stability of the materials created, serving in particular as a basis for the training of the laboratory’s AI.

The device is also focused on the handling and characterization of solid inorganic powders. In fact, solid powders are easily adaptable to production and technological increase. A-Lab’s synthesis approach produces sample ranges of varying quantities, facilitating testing at the intended device level. The results of the study are detailed in the journal

Autonomous discovery of new materials with A-Lab. © Nathan J. Szymanski and others.

Production of more than two new materials per day

O Materials Design is the world’s most widely used open access information repository (more than 400,000 users) for inorganic materials. Created in 2011 by Berkeley Lab, it brings together data on the properties of hundreds of thousands of known and predicted (yet non-existent) crystal structures and molecules. With the contribution of Google DeepMind, the database now has almost 380,000 additional new compounds (compared to a total of 154,718 before).

The new data provided by Google DeepMind was obtained through Graph Networks for Materials Exploration (GNoME), a deep learning tool trained with data previously provided by the Materials Project. The GNoME algorithm was then improved through active learning and generated 2.2 million new crystal structures. The 380,000 added to the Berkeley Lab database were selected for their stability and high potential for technological application.

Together, data from GNoME and the Materials Project helped establish a solid foundation for A-Lab. As part of the new study dedicated to the platform, all the materials targeted are entirely new, which means they are not present in the training data of the algorithm it uses to propose synthesis recipes. Furthermore, because the laboratory processes samples exposed to open air, the targets considered are not expected to react with O.two (oxygen), COtwo (carbon dioxide) and HtwoOh (water).

In just 17 days, A-Lab managed to generate 41 new compounds out of the 58 previously targeted, that is, a frequency of more than two new compounds per day. For comparison, it would take a human researcher several months to predict the structure of a material and experiment with it to determine whether it is viable or not. This success rate (71%) could be improved further, according to the study experts.

We show that combining theory and data with automation produces incredible results “, explain Gerbrand Ceder, principal investigator at A-Lab. “ We can manufacture and test materials faster than ever before, and adding more data points to Materials Design means we can make even smarter choices ” he adds.

Perspectives for use in sustainable technologies

With potential material innovations, it would be possible to produce new fully recyclable and biodegradable plastics, take advantage of wasted thermal energy with solar panels, design better batteries or more efficient and sustainable fuels, more efficient transistors for computers, etc.

We need to create new materials if we are to face global environmental and climate challenges said Kristin Persson, founder and director of the Materials Project, professor at the University of California, Berkeley, and co-author of the new research.

Previously, data from the Materials Project allowed experimental testing of various materials. These include, for example, carbon sensors, photocatalysts (materials that speed up chemical reactions in response to light and can be used to break down pollutants or generate hydrogen), thermoelectric materials (converting heat into electricity), and transparent conductors (used in solar cells). and LED and touch screens).

However, materials discovery is just one of many steps leading to their real-world application. For example, it takes a long time to empirically determine whether they have the right properties to match targeted technologies or whether they are cost-effective enough for mass production. However, A-Lab can significantly accelerate these steps and will ultimately provide a greater number of viable target materials for companies and industries.

Lab presentation video:

Source: Nature

Leave a Comment