Artificial intelligence inverse designs smart materials

Schematic representation of some of the quasicrystals designed by the AI system
Schematic representation of some of the quasicrystals designed by the AI system

Utrecht researchers have developed an artificial intelligence (AI) system that takes an important step towards designing new materials more efficiently. Based on the desired characteristics of a material, the AI system determines which building blocks are needed to make such a material. The researchers publish the results today in Science Advances.

Smart materials, such as catalysts or membranes, consist of atoms, molecules or nanoparticles that arrange themselves into a particular structure. “Normally, new materials are found by trial and error: out of an infinite number of possible atoms, molecules and compounds we try to find with painstaking measurements which chemical compositions lead to materials with interesting new properties,” says research leader Prof Marjolein Dijkstra. “But because there is an astronomical number of different atoms, molecules, and particles, it is impossible to navigate through all possible combinations.”

The new AI system reverses the process. “We first determine what kind of structure or material we want. The AI system then calculates  which molecules, compounds or particles are needed, and under which conditions, such as pressure, density and temperature, the material is stable.” The researchers use an evolutionary algorithm based on natural selection to adapt and optimise the building blocks and conditions to obtain the desired structure. An artificial neural network then assesses if the desired structure has been found. If not, the building blocks and conditions are further adapted in the next generation of the evolutionary algorithm.

CO2  capture and storage

In the publication, the researchers show that the AI system successfully inverse designs  quasicrystals. A quasicrystal is a structure that is ordered but not periodic. Quasicrystals have applications in, for example, photonics. Dijkstra envisions many future applications: “Next, I will use this AI system to design materials for a sustainable future, for example porous materials that are optimized to capture and store CO2.”

Emergence of artificial intelligence

Artificial intelligence is mainly known for applications such as search engines and self-driving cars, but it is now increasingly being used for scientific applications, explains Dijkstra. “More and more material scientists are working with AI these days. This is largely because computers have become faster, but also because we have access to large amounts of data that are yet to be explored. This offers new opportunities in the search for new materials. Using machine learning, we can efficiently explore the search space of an infinite number of material combinations.”


Inverse design of soft materials via a deep-learning-based evolutionary strategy
Gabriele M. Coli, Emanuele Boattini, Laura Filion, and Marjolein Dijkstra
Science Advances, 19 januari, DOI 10.1126/sciadv.abj6731

All authors are affiliated with Utrecht University