An approach that analyzes materials data with machine learning and other methods to help predict properties and choose which materials to try next. Combined with computation and experiment, it speeds up materials development.
Japanese name: マテリアルズ・インフォマティクス
Materials informatics appears in 1 practice question across 1 topic, including Computational materials science.
Materials informatics (MI) is a field that speeds up materials research and development with the methods of data science. Counting combinations of elements, compositions, structures and processing conditions, the candidate materials are enormous in number, and they cannot all be tried. So relationships between composition or structure and properties are learned from past experimental and computational data, and promising candidates are narrowed down. Spurred by the Materials Genome Initiative launched in the United States in 2011, the field spread worldwide, and research is advancing in universities and companies in Japan as well. Databases built by automatically running first-principles calculations on large numbers of materials also support this trend.
A typical flow is to collect data, build descriptors, train and evaluate a model, and use its predictions to choose the next experiments or calculations. Descriptors include the means and differences of the electronegativities and atomic radii of the constituent elements, the number of valence electrons and features of the crystal structure; choosing physically meaningful descriptors tends to give good models even with little data. Models include linear regression, random forests, Gaussian process regression and neural networks, and graph neural networks that treat crystal structures directly as graphs are also used. Examining which descriptors are effective can also give clues to the factors that govern a property.
Materials data are biased in that successful results tend to be published while failed ones are rarely recorded. Measurement conditions and specimen preparation also vary between papers, so values for the same material often scatter. In addition, if materials of very similar composition are in both the training and validation sets, the validation error comes out small and the model's ability on new families of materials is overestimated (data leakage). To avoid this, data are split and validated in groups by compositional family or type of compound. When predictions miss, it is also important to suspect the quality of the data rather than the model.
In inverse design, which searches for materials from properties, generative models and optimization methods are used to propose compositions and structures with the desired properties. Research is also advancing on autonomous experiments, in which active learning, alternating prediction and experiment, is combined with robotic synthesis and measurement and run with almost no human intervention. Also in use are process informatics, which learns relationships among processing conditions such as temperature and time, microstructure and properties, and surrogate models, which replace time-consuming simulations with machine-learning models to speed them up. Combining computation, experiment and data is the hallmark of this field.
Answer: B. Performance should be judged by the error on data not used for training (e.g. by cross-validation), and predictions far outside the training data's range are unreliable
A machine learning model learns a function that predicts a target property from descriptors (features) that express composition, crystal structure and so on as numbers. If it is fitted too closely to the training data (overfitting), the error on that data is small but predictions for new data are poor. The data are therefore split into training and validation sets, or cross-validation is used, and performance is evaluated by the error on data not used for training. Adding too many descriptors when data are scarce makes overfitting more likely. Models are also good at interpolating within the range of the training data but poor at extrapolating far outside it. For this reason, data are increased efficiently by methods such as Bayesian optimization, which uses both the predicted value and its uncertainty to choose the next candidate to try.
Computational materials science finds properties by solving equations based on physical laws. Materials informatics learns relationships statistically from data obtained by computation and experiment. The two are not opposed; they are combined, for example by generating data with first-principles calculations and searching widely with machine learning.
Approaches include using a small number of physically meaningful descriptors, using transfer learning to repurpose a model trained on a large dataset for a similar property, and adding computed values as auxiliary data. With little data, Gaussian process regression, which can also show the uncertainty of predictions, is often used.
It is a representation of the features of a material's composition or structure as a sequence of numbers, so that the material can be input to a machine-learning model; it is also called a feature. Means of elemental properties, crystal symmetry and the like are used, and choosing descriptors physically related to the property to be predicted improves both the accuracy and the interpretability of the model.
CALPHAD method Density functional theory Finite element method Interatomic potential Molecular dynamics Monte Carlo method Phase-field method