When we talk about Artificial Intelligence, we often think of chatbots or the possibility of automating some everyday tasks. In reality, one of the fields in which this new technology is finding applications is scientific research in a sector that is perhaps less well known: materials research.
When we talk about “discovering a new material,” we might think that it simply means finding a substance that has never been used before. In the field of research, however, the concept is more complex, because a material is not only defined by the chemical elements that make it up, but also by the way in which the atoms are organized, the properties that result from them, and the structure they take.
So, in simpler terms, it could be said that the search for new materials consists of looking for new combinations of elements and structures that may have interesting characteristics; these could include heat resistance, magnetic properties, the ability to store energy, or other characteristics that are useful for a specific application.
This type of research is important for technological development; for example, batteries, electronic devices, semiconductors, and many other technologies depend on the properties of the materials used to manufacture them.
The problem that may arise is that there are many chemical elements that can be combined, and they can be organized according to different structures. Not all the combinations that are conceived or studied are, in practice, feasible: some structures are unstable, while others that may be technically interesting could be difficult to produce in the laboratory.
Traditionally, the search for a new material requires several steps: it may start with a hypothesis or with a property that one would like to obtain; calculations and simulations are then carried out to understand its actual stability; finally, if the results are promising, the experimental phase begins, attempting to actually produce the material and verify its characteristics.
And it is precisely in this process that Artificial Intelligence can intervene.
Machine learning systems can be trained using large amounts of data relating to materials that have already been studied. Based on this, they could learn to recognize certain relationships between the composition, structure, and properties of materials, which can then be used to identify new candidates for further analysis.
The goal is therefore not for AI to “invent” a new substance, but rather for it to help researchers explore a very large number of possibilities, identifying those that might be more interesting.
One of the best-known examples is GNoME, a system developed by researchers at Google DeepMind and described in 2023 in the journal Nature. The model was designed to analyze crystal structures and predict their stability.
In their study, the researchers reported the identification of 2.2 million crystal structures that, according to the criteria used, were new and potentially stable. This, however, does not mean that 2.2 million new materials were produced in the laboratory; rather, these are structures identified through a computational process that can become candidates for further in-depth studies. A prediction obtained by a computer does not automatically correspond to the experimental discovery of a material.
More recent research is attempting to make this process even more effective. In 2025, for example, a study on MatterGen was published in Nature, focusing on a generative Artificial Intelligence model designed to create structures of inorganic materials and guide them toward specific characteristics. In this case, AI does not simply select from materials already present in the available data, but can generate new candidate structures based on what it has learned. The researchers also reported the experimental synthesis of one of the structures generated by the model, as evidence that—at least some of the predictions—can be taken from simulation to the laboratory.
Here too, however, it is important not to confuse a prediction with a result that is already ready to be used. Experiments and verification are therefore necessary in order to compare the model’s predictions with what actually happens in reality.
It is precisely in the integration of these different steps that Artificial Intelligence can play an important role, identifying promising candidates and contributing to understanding which experiments should be carried out. The new data obtained will then be useful for providing additional data that can be used to guide subsequent research. In this way, simulation and Artificial Intelligence can become part of the same process.
AI, therefore, does not replace the researcher and does not turn a prediction into a discovery, but its contribution lies in the possibility of exploring a very large number of possibilities and making research more efficient. It is the collaboration between computational capabilities and experimental verification that could, in the future, make it possible to identify materials with useful properties for technology.
Author: Ms Maria Lagani, psychology graduate- Master’s student. Research Team for JUMP staff (Italy)
Source: Merchant, A., Batzner, S., Schoenholz, S. S., Aykol, M., Cheon, G., Cubuk, E. D., et al. (2023). Scaling deep learning for materials discovery. Nature, 624, 80–85. https://doi.org/10.1038/s41586-023-06735-9
Szymanski, N. J., Bartel, C. J., Zeng, Y., Tu, Q., Ceder, G., & others (2023). An autonomous laboratory for the accelerated synthesis of inorganic materials. Nature, 624, 86–91. https://doi.org/10.1038/s41586-023-06734-w
Zeni, C., Pinsler, R., Zügner, D., Fowler, A., Horton, M., Fu, X., et al. (2025). A generative model for inorganic materials design. Nature, 639, 624–632. https://doi.org/10.1038/s41586-025-08628-5