Materials are an essential part of our world; they have enabled us to build cities, treat disease, and communicate across the world in real time. For centuries, material scientists have been working to build our material library and to discover new materials with greater performance and better property trade-offs. Over the past few decades, however, the rate at which new materials are being discovered has been slowing continuously. This is due to several factors including increasingly stricter regulations, more challenging performance metrics, and increasingly more expensive empirical development strategies.
The reduced rate of material discovery is also in part because many of the simplest combinations have been investigated, and the number of remaining possible combinations is quite extensive. For example, if random combinations of five elements from the periodic table are combined in equal amounts, there would be 1078 possible combinations to choose from. This example ignores the fact that different numbers of elements can be combined, not just five, and that they do not have to be combined in equal ratios. For perspective, there are estimated to be 1066 atoms in the Milky Way galaxy. So, in terms of a “big data” challenge, materials development of complex composition alloys represents perhaps the biggest big data paradigm. There is a clear need for a method to narrow the search space to only the most promising candidates for a given application. Intuition and expensive trial and error strategies will not be sufficient for investigating this immense chemical space, and more informed computational methods must be developed and employed.
Our team of nanoengineers in Professor Kenneth Vecchio’s lab at UC San Diego is developing tools for screening large numbers of materials in a rapid fashion. The first step in our work is creating unique identifiers for each material, akin to a fingerprint of the material. In the same way no two fingerprints are alike, every individual material possible can be reduced to a simple but unique set of attributes. These identifiers describe the material composition in a way that supports computation work leveraging a subset of artificial intelligence called machine learning. The machine learning tools learn the underlying science that relates these attributes to various material properties. Typically, machine learning requires enormous initial datasets to learn from before it becomes a useful tool.
However, the method that our team developed is designed around the fact that material development problems frequently have less than 100 data points at the outset. After learning about the initial dataset, the machine learning algorithm suggests new materials with the goal of maximizing performance. Each time the materials suggested by the algorithm are fabricated and tested, this new information is made available to the algorithm, creating a learning loop.
The data-driven method that our team has developed was recently demonstrated for predicting the synthesizability of single crystal structure (e.g. rock-salt structured) carbide ceramic materials containing five metal cations, also known as high entropy carbides. High entropy carbides constitute a subset of the complex concentrated alloys class of materials described previously, as they have the added uniqueness of becoming more stable at increasing temperatures, which is unlike most engineering materials. The researchers focused their study on what are called non-intuitive compositions, in which three of the five metal cations are chromium, molybdenum, and tungsten, none of which form a rock-salt structure at room temperature in a one metal atom to one carbon atom ratio.
The initial dataset contained all available data: 56 high entropy carbide materials with synthesizability calculated by computationally expensive density functional theory (DFT). None of the 56 known compositions contained chromium, one of the three metal cations of interest. While DFT can compute a few compositions per month, the machine learning model was able to learn from the 56 materials and make predictions on 70 new materials in less than one day.
Seven materials, four predicted to succeed and three predicted to fail, were experimentally fabricated and analyzed to assess the validity of the predictions. Rather surprisingly, several five-cation metal carbide compositions were discovered, wherein three of the five cations were chromium, molybdenum, and tungsten—the elements that don’t form the rocksalt monocarbide structure—and yet these compositions were experimentally shown to successfully form the rock-salt structure. Furthermore, all seven experimentally studied compositions resulted in single or multi-crystal structure materials in exact agreement with the machine learning predictions. The ability for the machine learning model to perform exceedingly well in such a non-intuitive chemical space, a composition space which contained no prior data to learn from, further demonstrates the unique strength of this approach. Our team expects the machine learning framework to be a useful tool in the development of other materials such as alloys, battery components, or pharmaceuticals.
This work is published in Nature Partner Journals (npj) Computational Materials, May 1, 2020.
Read the paper here: https://rdcu.be/b3UdG
DOI: 10.1038/s41524-020-0317-6
