📖 MI Glossary

A research field that combines materials science with data science and machine learning to discover and design materials with desired properties more efficiently. It replaces the traditional trial-and-error, intuition-driven approach with one guided by data and algorithms. The aim is to dramatically shorten the time needed to find new materials.

Predicting what properties a given structure or composition will have is the forward problem; working backward from a target property to the material that produces it is the inverse problem. The ultimate goal of MI is to solve the inverse problem, that is, to design a material starting from the performance you want. Inverse design is much harder because many candidates may lead to the same property.

An approach that starts from data collected through experiments, simulations, and the literature, then uses statistics and machine learning to drive materials development. Rather than relying only on theory or empirical rules, it extracts trends and patterns from the accumulated data itself. This is the central mindset of MI.

A strategy that evaluates a huge number of candidate materials at once, using computation or automated experiments, to narrow down the promising ones. It can sift through tens or hundreds of thousands of candidates that would be impossible to handle by hand. Combining it with machine learning predictions makes the screening even more efficient.

The enormous space formed by every possible combination of composition, structure, and processing conditions to be explored. Treating element types, ratios, and crystal structures as variables leads to an astronomical number of combinations. Efficiently exploring this vast space is a central challenge of MI.

The correspondence between a material's "structure" (its atomic arrangement and composition) and its "properties" (strength, conductivity, catalytic activity, and so on). If this relationship can be understood and modeled, properties can be predicted from structure, or structures designed from target properties. It is a foundational concept in both materials science and MI.

A U.S. national program launched in 2011 that aimed to halve the time and cost of developing new materials. It popularized the idea of tightly linking computation, experiment, and data infrastructure, and became a starting point for MI efforts worldwide. Many countries, including Japan, have since developed similar national strategies.

A technique that arranges many materials on a map according to their properties or features, making the overall trends and empty regions easy to see. Dimensionality reduction is used to project high-dimensional data into two or three dimensions and reveal unexplored, promising areas. It helps in planning exploration strategies.

An empirical rule in which one physical quantity shows a regular proportional or power-law relationship with another. In catalysis, for example, the adsorption energies of different reaction intermediates are known to fall on straight lines, which guides materials design. Such relations let a simple indicator stand in for a more complex performance metric.

A material converted into a vector of numbers that a machine learning model can handle. By turning composition, structure, or electronic-state information into numbers, a model can compare materials and predict properties. Designing good descriptors is a decisive step that strongly affects MI accuracy.

Features computed from a material's chemical composition alone, that is, the elements it contains and their ratios. They are built by statistically aggregating element properties such as atomic mass, electronegativity, and valence electron count. Because they can be computed even when the crystal structure is unknown, they are widely used in early screening.

Features that reflect the three-dimensional arrangement of atoms and the nature of their bonds in a crystal or molecule. They carry more information than composition-based features and tend to give higher accuracy, but they require structural data. This representation pairs naturally with graph neural networks.

A widely used feature library that automatically generates a standard set of descriptors from composition. It computes statistics of each element's physical and chemical properties, weighted by composition ratio. Easily accessible through tools such as matminer, it has become a de facto standard for composition-based MI.

A structural descriptor that represents the local environment around an atom using smooth functions. It is invariant to rotation and translation and captures subtle differences in atomic arrangement continuously. It is widely used in machine learning potentials and in computing structural similarity.

A descriptor that encodes whether certain substructures are present in a molecule as a string of zeros and ones. It allows the similarity between molecules to be computed quickly, which is useful for compound search and screening. It is one of the basic molecular representations in cheminformatics.

A method that turns categorical information, such as an element name or crystal system, into a vector where only the matching entry is one and the rest are zero. It represents categories without implying any order or magnitude among them. It is simple, but the number of dimensions grows quickly when there are many categories.

The work of creating or selecting useful features from raw data so that a model can learn from them more easily. Designing good features with domain knowledge can yield high accuracy even from limited data. In MI it is, alongside model choice, one of the steps that most determines success.

A technique that compresses many features (high-dimensional data) into a small number of variables while preserving as much information as possible. Methods such as PCA, t-SNE, and UMAP are common and help with visualization and noise removal. They are used when viewing the big picture of a materials space as a two-dimensional map.

The work of removing errors, duplicates, outliers, and unit mismatches to put data into a usable state. Experimental and literature data are often incomplete, so this preprocessing strongly affects prediction accuracy. It is the basic step for avoiding the "garbage in, garbage out" problem.

Places where a value is absent because part of the data was never measured or the record is missing. Many machine learning models cannot use such data as-is, so the values are either removed or filled in (imputed) with a mean or a model-based estimate. Because the choice of handling affects results, it calls for careful judgment.

Preprocessing that puts features on a common scale, because differing units and ranges can bias learning. Standardization transforms data to zero mean and unit variance, while normalization squeezes values into a range such as 0 to 1. It is an especially important step for models that rely on distances or gradients.

A method that learns a model to predict an output from an input by being given pairs of inputs and correct labels (for example, composition and a measured property). Most property prediction in materials falls under supervised learning. The key is whether enough labeled data can be prepared.

A method that finds structure or groups hidden in data without using correct labels. Clustering and dimensionality reduction are typical examples, used to classify materials into types or uncover hidden trends. It is helpful in the early stages of exploration when labeling is difficult.

A supervised learning task that predicts a continuous numerical value, such as melting point, band gap, or strength. It approximates the relationship between inputs and outputs as a function. Because many material properties are continuous quantities, regression is the most common task in MI.

A supervised learning task that assigns data to one of several predefined categories. Predictions such as "metal or insulator" or "stable or unstable" fall into this category. The difference from regression is that the output is a discrete label.

A state in which a model fits the training data too closely and can no longer predict well on unseen data. It happens when the model memorizes noise rather than the essential trend in the data. It is especially common in MI, where data is scarce, and is countered with cross-validation and regularization.

A technique that splits data into several groups and repeatedly trains and evaluates while holding out part for validation, to estimate a model's performance. It makes effective use of limited data and helps detect overfitting. With small datasets common in MI, it is a standard evaluation procedure.

The ability of a model to predict correctly on unseen data that was not used for training. Because the goal of MI is to predict unknown materials rather than reproduce known data, generalization is the most important measure. It is judged by validation error, not training error.

A value set by a human before training that governs the model's structure or how learning proceeds, such as tree depth or learning rate. It is distinct from the parameters that are learned automatically from data. Tuning it appropriately can greatly improve performance.

A model that repeatedly asks yes-or-no questions ("is this feature greater than some value?") and branches until it reaches a prediction. Because it expresses the reasoning as a tree, it is easy to interpret why a given prediction was made. It can be used on its own, and also serves as the building block for random forests.

An ensemble method that builds many decision trees under slightly different conditions and combines their predictions by voting or averaging. It is more accurate and stable than a single tree and relatively resistant to overfitting. Easy to use even with small datasets, it is a go-to model in MI.

A model of many "nodes" connected in layers, inspired by the brain's neural circuits. It transforms the input step by step through the layers and can express complex, nonlinear relationships. Many modern MI methods, including deep learning and GNNs, are built on it.

A method using neural networks with many stacked layers that automatically extract features from data as they learn. It performs strongly on complex data such as images, language, and molecular structures. It shows its full power when enough data and computing resources are available.

A neural network that represents a material as a graph, with atoms as nodes and bonds as edges, and learns on that graph. Because it naturally handles crystal and molecular structure, it achieves high accuracy in property prediction. Its strength is learning directly from structure without hand-designed features.

An architecture, foundational to large-scale models, that uses an "attention" mechanism to capture the important relationships within data. Originally introduced for natural language processing, it is now applied to representation learning for molecules and crystals. It is the core technology behind foundation models pre-trained on large amounts of data.

A function that expresses numerically how far a model's prediction is from the correct answer. Learning is the process of adjusting parameters to make this loss as small as possible. Depending on the task, one chooses mean squared error for regression, cross-entropy for classification, and so on.

An efficient optimization method for finding the best conditions in as few trials as possible. It builds a surrogate model from past results and cleverly decides "where to try next." It pairs extremely well with materials exploration, where each experiment or calculation is costly.

The criterion in Bayesian optimization for deciding the next point to try. It combines how high the predicted value is (exploitation) with how large the uncertainty is (exploration) to pick the most valuable candidate. Several kinds exist, such as EI and UCB, and they shape the character of the search.

A lightweight model that approximates the input-output relationship in place of a real experiment or a heavy calculation. Because it can evaluate many candidates cheaply, the optimization loop can run quickly. In Bayesian optimization, a Gaussian process is a common surrogate model.

A regression model that outputs not only a predicted value but also how confident it is (the uncertainty) in that prediction. Because uncertainty grows in regions with little data, it helps decide where to explore. It plays a central role as the surrogate model in Bayesian optimization.

The dilemma of whether to prioritize investigating untried regions ("exploration") or digging deeper into promising ones ("exploitation"). Leaning too far toward exploration is inefficient, while leaning too far toward exploitation risks getting stuck in a local optimum. Balancing the two is the key to efficient optimization.

A framework in which the model itself chooses "which data to acquire next to learn the most," proposing experiments or calculations. By gathering the most informative data first, it can build an accurate model in few trials. It is powerful in materials development, where experiments are expensive.

A problem of optimizing several competing goals at once, such as "high strength and low weight." There is usually no solution that perfectly satisfies all of them, so one seeks a set of well-balanced solutions. Real materials design tends to be a multi-objective problem that must reconcile many requirements.

In multi-objective optimization, the set of solutions at the boundary where improving one objective necessarily worsens another. Every solution on this front is an equally good compromise, and the designer chooses among them according to the goal. It is an important concept for visualizing trade-offs.

A framework in which an agent acts through trial and error and learns a policy that maximizes the reward it receives. It is applied to materials exploration that requires a sequence of decisions, such as designing synthesis routes or choosing experimental conditions step by step. A distinctive feature is that it can learn even without explicit labeled data.

A setup that automatically cycles through an AI proposing the next experiment, a robot carrying it out, and the resulting data updating the AI. It accelerates materials exploration by keeping human intervention to a minimum. It is the core idea behind autonomous (self-driving) laboratories.

A statistical planning method for investigating the effects of factors efficiently with as few experiments as possible. Using orthogonal arrays or response surface methods, it obtains information more efficiently than experimenting at random. It is a classic framework for designing experiments cleverly, predating Bayesian optimization.

A representative first-principles method that computes a material's energy and electronic states based on the electron density distribution. Because it can predict properties without experiments, it is widely used to generate data for MI. Its good balance of accuracy and computational cost has made it the standard for materials calculations.

A method that computes a material's properties from the fundamental equations of quantum mechanics alone, without relying on empirical parameters. It can predict properties of even unknown materials before experiments, serving as an important data source for MI. It is computationally expensive but highly reliable.

A method that numerically solves Newton's equations of motion while computing the forces on atoms, reproducing atomic motion over time. It can study temperature- and time-dependent phenomena such as diffusion, thermal conduction, and phase transitions. Using a machine learning potential for the forces can speed it up dramatically.

A model that learns from first-principles data to predict energies and forces from atomic arrangements at high speed. It delivers DFT-level accuracy at the speed of a classical force field, enabling large-scale, long-timescale simulations. It has spread rapidly in computational materials science in recent years.

A multidimensional "landscape" showing how energy changes as the arrangement of atoms is varied. Valleys correspond to stable structures and passes to reaction barriers, forming the basis for understanding material stability and reaction pathways. A machine learning potential can be seen as an attempt to approximate this surface.

A quantized description of the collective vibrations of atoms in a crystal. It is essential for understanding thermal properties such as heat capacity, thermal conductivity, and thermal expansion, as well as structural stability. Computing phonons from first principles allows a material's thermal properties to be predicted.

A diagram showing how the energies electrons can occupy in a crystal are distributed with respect to momentum. The presence and size of a band gap reveal whether a material is a metal, semiconductor, or insulator, along with its optical properties. It is one of the most fundamental pieces of information in electronic materials design.

A distribution showing how many states are available for electrons at each energy. It is the band structure aggregated over momentum, and is used to discuss electrical conduction, magnetism, and catalytic activity. The abundance of states near a particular energy strongly influences properties.

The energy exchanged when a compound is formed from its constituent elements, serving as a measure of whether the material is thermodynamically stable. Lower values indicate greater stability and a higher likelihood that it can actually be synthesized. It is an important indicator of whether a candidate material could exist in reality.

The part of DFT that approximately represents the complex electron-electron interactions known as exchange and correlation. Many kinds exist, such as LDA, GGA, and hybrids, and the choice changes the calculation accuracy. It determines the accuracy and applicable range of DFT, in effect setting the "skill level" of the calculation.

A leading materials database that freely publishes DFT-computed property data for hundreds of thousands of inorganic materials. Composition, crystal structure, formation energy, band gap, and more can be retrieved through an API. It is used around the world as a starting point for MI research.

An open database holding a large amount of DFT-based thermodynamic and structural data for inorganic materials. It is especially rich in formation energy and phase stability data. Alongside Materials Project, it is a frequently used source of training data for machine learning models.

A framework and database that automates high-throughput first-principles calculations to generate and publish vast amounts of materials data. Millions of computed results can be searched and used. It is applied to the discovery of alloys and novel compounds.

A European repository that stores and shares computational materials data worldwide, including the original input and output files. It handles data from different software in a unified way, emphasizing reproducibility and data sharing. It is known as a data platform that puts the FAIR principles into practice.

An international standard that lets multiple materials databases be searched together through a common API specification. You can access several data sources with the same query style, without learning each database's separate query method. It is an effort to improve the interoperability of data.

The collective name for the family of materials databases operated by Japan's National Institute for Materials Science (NIMS). It provides diverse data, including experimental data on polymers, inorganics, metals, diffusion, and electronic structure. As a Japanese MI platform, it is widely referenced in domestic research.

A Python materials-analysis library for manipulating crystal structures, analyzing properties, and accessing databases. It integrates closely with Materials Project and makes reading, writing, and converting structure data easy. It has become the de facto standard tool for data processing in MI.

A Python library that automatically generates machine learning features from materials data and supports data retrieval and visualization. It can compute many descriptors, including Magpie, in just a few lines of code. It greatly streamlines the preprocessing for composition- and structure-based MI.

A Python toolkit for creating and manipulating atomic structures and controlling various calculation programs in a unified way. It can drive many calculation engines, such as DFT and molecular dynamics, with a common interface. It serves as a foundation for automating calculation workflows and for high-throughput computing.

A library for implementing graph neural networks on top of PyTorch. It treats molecules and crystals as graphs and lets you build and train GNN models efficiently. It is a standard tool, widely used in GNN research on materials and molecules.

Guidelines for making data Findable, Accessible, Interoperable, and Reusable. They aim to share data in a form that is easy for both humans and machines to use. They are a valued principle in building the data infrastructure for MI.

An open-source tool that connects algorithms such as Bayesian optimization to robotic experiments and calculations to run a closed loop for materials exploration. It automates everything from proposing "what to try next" to incorporating the results. It is drawing attention as a platform for easily getting started with autonomous materials development.