Machine learning for materials design
Exploring a patent for a machine learning approach to materials design.
Like in all fields of science and technology right now, there is an increasing use of machine learning based tools in materials science. One particular area that has been the focus of intensive research is materials discovery.
Traditionally, novel materials have been unearthed through real-world exploration and painstaking synthesis experimentation, and, sometimes, quite literally, by digging new mineral forms out of the ground. In the last few decades, computationally intensive electronic structure methods (such as density functional theory (DFT)) and other materials modelling techniques have been used to predict the structure and simulate the behaviour of new candidate materials.
More recently, however, attempts have been made to apply machine learning and other artificial intelligence (AI) models to databases of experimental and theoretical results to accelerate the process of materials discovery.
Many advances have already been made. For instance, in 2023, Google DeepMind announced that a deep-learning AI model GNoME (trained on DFT-predicted structures) had discovered 2.2 million new crystalline materials, of which 381,000 were considered the most thermodynamically stable and so promising for experimental synthesis.
Microsoft has also recently developed the MatterGen and MatterSim tools. MatterGen is a machine-learning model designed to predict potential materials with desirable properties, while MatterSim evaluates which of the imagined materials may be stable and viable.
However, there are still many challenges facing AI-assisted materials design. One issue is that electronic structure calculations typically predict ordered crystal structures that are stable at absolute zero. These may not consider the important effects of temperature on structural transformations and disorder.
Another issue is the difficulty in identifying routes for synthesising and testing predicted materials. Further research is therefore required into models that more accurately predict realisable materials with desirable properties for given applications.
European patent EP3800586B1, granted in January 2025 to Samsung Electronics Co, Ltd., claims protection for a new machine-learning method for materials discovery that seeks
to address at least some of these concerns.
According to the patent, traditional materials design typically follows a conventional approach described as the “forward” mode. In this mode, a physical property (P) of interest (e.g., resistivity, melting point, or oxidation resistance) is identified.
Candidate materials (M) are then gathered based on prior knowledge or chemical intuition, and similarity with known materials having property P. Suitable potential material structures (S) are chosen (such as the crystal structure or orientation), and property P is calculated for each structure using continuum or atomistic simulations.
A decision is then made about whether each material M with structure S meets the target property P and so is a favourable candidate.
The inventors outline two key problems associated with this mode of design. First, the selection of candidate materials is based on prior knowledge and intuition. This means that potentially suitable target materials may be excluded, because they lie outside the intuition of experts in the field or are too dissimilar to known materials having the desired property.
Second, the simulation of property P for various structures is a time-consuming process, meaning that only a minority of candidate materials are realistically considered.
The patent proposes a machine-learning system based on a generative approach to “inverse design”.
In inverse design, the target property P is used as the desired performance criterion in an optimisation problem to find suitable candidate structures S automatically.
The patent outlines three different approaches to materials inverse design – a statistical inference-based approach, a genetic algorithm-based approach and a generative model-based approach, as followed in the patent.
The patented method involves a generative adversarial network (GAN). This comprises a discriminator network that learns the underlying joint probability distribution p(S,P) of structure S, with the property P relationships for real structures and real properties recorded in a database simultaneously (i.e., not sequentially). The method uses this learning to directly generate a large number of candidate samples (S,P) and filter those that meet the target property P.
This generative method moves beyond the traditional approach of selecting from a limited list of candidate materials, or relying on human intuition, potentially enabling the discovery of useful candidates having unexpected structures.
By directly learning the intrinsic relationship between a material’s structure and its properties, as represented by the joint probability distribution p(S,P), the method avoids the need to calculate properties after a candidate structure has been chosen. This approach promises to be faster than those requiring sequential simulation steps.
According to the patent, a discriminator network that is trained to distinguish simulated structures and properties from real structures and properties (and thereby generate a loss function that is fed back to a generator network), forces the system to produce more realistic, and thus credible, samples.
