First-principles computations have been widely adopted in the last two decades as a tool to design and characterize new materials and molecules. For a known compound, i.e., when the crystal or molecular structure is known, methodologies such as density functional theory (DFT) allow to quantify the intrinsic properties (e.g., electronic structure, transport of heat, electrons, or ions, light absorption and emission, hydrogen and oxygen evolution reaction energies, etc.) and determine whether it is suitable for a given application.
Over the last 20 years, various databases of such computations have been constructed and made available to researchers. These can be used to scan the existing compound landscape to search for adequate materials and molecules. More recently and with the advent of machine learning, these databases have been used to train interatomic potentials applicable to the whole periodic table. These so-called universal machine-learning interatomic potentials allow DFT-precision simulations on a scale 3 orders of magnitude larger than what is reachable through classical DFT, reducing the gap with the technology scale and opening new possibilities and research routes. All these approaches have been employed in NICKEFFECT for the design of magnetic Ni-Co alloys for RAM applications.
However great these approaches are to materials design, they are limited to known crystal structures and molecules. Beyond investigating already-synthesized materials and molecules, until recently, designing new compounds has relied on element substitutions and permutations in known structure families. Even within such constraints, the large dimensionality of the search space is such that systematic optimization is intractable. Recently, generative artificial intelligence (GenAI) approaches have been developed and used to generate new materials beyond those included in the training database. GenAI encodes the structures (molecules or crystals) from a database into a smaller-dimension latent space. Using the knowledge embedded in this latent space, new structures may be easily suggested and further refined using traditional computational methods based on ML or first principles.
This two-step approach, where the crystals or molecules are first generated using GenAI and then screened using computational methods (eventually followed by experimental characterization and validation), still leads to significant computational (and experimental) effort before finding an appropriate candidate. Alternatively, latest approaches progressively include those additional characterizations as co-modalities of the GenAI model by aligning different latent spaces obtained for each target property. Such an approach would provide suggestions of compounds that are also aligned with the objectives in terms of physical and chemical properties, as well as synthesis parameters specific for each compound.
This new avenue for materials design is being developed by Matgenix and will be used in future projects beyond NICKEFFECT.