Materials & AI · glossary
Met an unfamiliar term while reading? Start with the concept.
39 terms
See allThe examples explain the concept. For actual research results, see AI in materials에서 원문과 함께 확인하세요. 엄선한 용어 86개에는 Colab에서 바로 돌려볼 수 있는 실습 노트북이 있습니다. “AI 자동 추가” 표시가 붙은 용어는 새 이슈를 수집할 때 자동으로 추가된 것으로, 아직 사람이 검토하지 않았고 실습 노트북이 없습니다. 노트북의 데이터는 개념 확인용으로 생성한 값이며 실제 측정값이 아닙니다.
Materials Informatics · MI · 소재정보학
The field that connects materials data with statistics, computation and machine learning to understand and design materials.
Gathering experimental and computed data to analyse composition-property relationships.
Composition Descriptor · Magpie · 원소 특성 통계
A representation used as model input, built from statistics of elemental properties that can be computed from the composition alone.
Using the mean and the spread of the constituent elements’ electronegativity as input.
Crystal Graph · CGCNN · 구조 그래프
A representation that turns a crystal structure into a graph of atom nodes and neighbour edges for a model to consume.
Connecting neighbours within a cutoff radius as edges to express the structure.
Machine Learning Interatomic Potential · MLIP · MLFF · 기계학습 포텐셜
A model that learns from energy and force data, typically from DFT, to predict energies and forces for an atomic arrangement quickly. It makes large and long simulations possible, but grows unreliable outside the range it was trained on.
Computing diffusion in a system of thousands of atoms with a trained potential.
Inverse Design
The approach that runs opposite to predicting a property from a composition: starting from the desired property and searching for a structure or composition.
Generating candidates that would meet a target permittivity.
Active Learning · AL
A method where the model chooses which data to label next and updates itself on the new result.
Computing the materials whose predictions are least certain and adding them to the training set.
Bayesian Optimization · BO
A method that uses predictions and their uncertainty to find good settings for an objective that is expensive to evaluate.
Choosing synthesis temperature and composition while spending as few experiments as possible.
Acquisition Function · EI · UCB · 기대 개선
A function that combines the prediction and its uncertainty into one score to decide which condition to evaluate next.
Picking the composition with the largest expected improvement as the next experiment.
Surrogate Model
A model that approximates the result of an expensive experiment or calculation quickly.
Narrowing candidates with a property model instead of running DFT every time.
High-throughput Screening · HTS · 다단 스크리닝
A staged search that filters many candidates cheaply and applies expensive verification only to what survives.
Cutting ten thousand candidates to a hundred with a predictive model before computing and measuring.
Closed-loop Discovery · 순환 탐색
A search that repeats proposing candidates, running the experiment or calculation, analysing the result and updating the model.
Changing the next composition proposal according to the newest experimental result.
Self-driving Laboratory · SDL · 자율실험
A system that combines automated equipment with data-driven decisions to choose and carry out the next experiment.
Having AI analyse a robot synthesis result and set the conditions for the next run.
Literature Mining · Text Mining · 문헌 마이닝
Extracting composition, process and property values automatically from the sentences of papers and patents to build data. The extracted values need checking against the source.
Normalising the ways synthesis temperature is written into standard units to build a dataset.
Matched Molecular Pair Analysis · MMPA
A method that pairs two molecules that differ only in one part and compares them to interpret how that small structural change affects properties or activity. It is useful for locally understanding structure-property relationships.
It can be used to compare how much the charge mobility changes when one substituent is changed in a candidate material molecule.
Demand Response
It is a method of cooperating with power grid operation by reducing electricity use or shifting the timing of use when power demand is high or the grid is unstable. It is a concept for flexibly managing load in energy systems.
The concept of Demand Response is important in data center operation, where large computing resources are adjusted to match power grid conditions.
Hierarchical Machine Learning
A method of learning by dividing data or prediction problems into multiple levels of hierarchy. The upper level can be designed to handle coarse structure, while the lower level handles fine variations.
The overall trend of a catalytic surface and the differences among individual active sites can be modeled together with Hierarchical Machine Learning.
Simulation Framework
An integrated structure that combines phenomena, inputs, state changes, and output evaluation into a single procedure to perform reproducible simulations. It systematically analyzes complex systems by linking physical models, data processing, inference, and validation steps.
It is used when constructing a computation pipeline that jointly interprets surface catalyst structures and spectroscopic signals.
Machine Learning Molecular Dynamics · MLMD
A method that rapidly performs molecular dynamics simulations by approximating interatomic interactions or forces with machine learning. It can more efficiently handle larger systems and longer time scales than expensive quantum-mechanical calculations.
It is used when quickly evaluating the diffusion of electrolyte molecules and the solvation structure.
Physics Constraint
A restriction imposed so that the solution a model predicts does not violate physical laws, symmetry, conservation laws, or the like. It is often used to improve training stability and generalization performance.
In crystal structure reconstruction, interatomic distances or sample symmetry can be incorporated as a physics constraint.
Motif Language
A structural representation method for expressing and manipulating complex structures as motifs, which are repeatable semantic units. It increases interpretability and editability, linking generation and design.
It is used when generating new structures by representing a crystal structure as a combination of motifs instead of at the atomic level.
Representation Learning
A method that automatically learns feature representations useful for downstream tasks from raw data. It focuses on extracting the structure and patterns of data without manual feature engineering.
It learns latent representations from raw crystal structure inputs that are advantageous for property prediction.
Local Electronic Descriptor
An expression that summarizes and quantifies electronic-structure information in the local environment around an atom. It makes microscopic electronic properties such as the bonding state and charge distribution around an atom comparable.
In materials inverse design, local electronic descriptors around atoms can be generated together to reflect material properties.
Machine-learning-assisted Molecular Dynamics · ML-assisted MD
A molecular dynamics technique that uses a machine-learning model to accelerate or correct calculations of interatomic forces or energies. It is used to handle larger systems or longer time scales than conventional MD.
Machine-learning-assisted Molecular Dynamics · ML-assisted MD can be applied to calculate ion migration barriers at a solid electrolyte interface.
Fragment-based Representation Learning
A method for learning representations by dividing complex targets such as molecules, polymers, or compositions into smaller fragment units. Rather than handling the whole object directly, it summarizes properties through combinations of fragments to support generalization and interpretability.
The repeating units and substituent information in a polymer composition are split into fragments and used as input for property prediction.
Image Restoration
A process that reduces noise, blur, distortion, and other degradations in a deteriorated image so that the original structure is revealed more clearly. It is used to improve data damaged by observation limits or imaging conditions through post-processing.
In a microscope image, noise is reduced to restore fine structures more clearly.
Intelligent Detector
A detector or detector system that analyzes data generated during measurement in real time and adjusts the next measurement conditions or data acquisition method. It is used to improve measurement efficiency and data quality in automated experimental environments.
In X-ray scattering experiments, an intelligent detector can immediately detect abnormal signals and adjust the measurement time.
Closed-loop Framework
A procedure that links the stages of experiment, analysis, and decision-making into a single cyclical process and automatically updates the next experiment. It improves efficiency by immediately reflecting the results obtained in each iteration in the next search.
In perovskite solar cells, candidate molecule screening and process optimization can be integrated into a single Closed-loop Framework.
Physics-Guided Active Learning
This is active learning that incorporates physical laws or domain knowledge into the learning strategy. It uses both the model's predictions and physical constraints to more efficiently choose the next candidates to experiment on or calculate.
Physics-Guided Active Learning can be designed to prioritize experimental evaluation of composite candidates with large interfacial barriers.
Substructure Model
A model that decomposes and represents a molecule, polymer, or crystal structure into smaller structural units called substructures for prediction. It has the advantage that the contribution of each partial structure to properties is easy to interpret.
Polymer dielectric candidates can be screened by dividing them into substructures such as repeat units and functional groups.
Quality Control
A procedure for inspecting and managing whether a product or specimen satisfies a specified standard. It includes measured values, defect detection, and pass/fail judgment.
After thin-film growth, thickness and defect distribution are checked as quality control metrics.
Process Control
A method for monitoring and adjusting the inputs and state of a manufacturing or experimental process to maintain the desired output. It can automatically correct conditions using sensor signals or model predictions.
In 2D materials CVD, temperature and gas flow are adjusted by process control.
Physics-aware Machine Learning
A machine learning method that incorporates physical laws, symmetries, conservation quantities, and constraints into the learning process or objective function. It helps enable predictions and generation that are physically valid, not just data-fitting.
In microstructure design, interfacial and crystallographic constraints are added to the loss function to guide feasible structures.
Latent Space Sampling
A method that selects points in a learned latent space and uses a decoder or generative model to produce actual candidates. It can efficiently generate new structures without directly handling the search space.
In microstructure inverse design, latent space sampling generates candidate structures outside the training data.
Prototype Bias
A phenomenon in which a model’s predictions or generation rely excessively on already known representative structures or templates rather than the new distribution itself. Even when apparent performance is high, actual generalization can be limited.
In generative crystal structure prediction, it is revealed as a sharp performance drop when existing prototypes are removed.
Physics-Informed Machine Learning · PIML
A machine learning method that incorporates physical laws, conservation equations, boundary conditions, and empirical constraints into the learning process or model architecture. Compared with learning from data alone, it can yield physically valid predictions and better generalization.
In catalyst screening, mechanistic constraints can be added to the loss function to reduce unrealistic candidates.
Semantic Learning
An approach that learns the conceptual meaning and relationships of data rather than its surface form. It is useful for understanding correspondences and context across different expressions.
In electrocatalysis literature, it groups expressions with similar functions and connects them to the same candidate set.
Information Extraction · IE
A method that finds desired facts in unstructured text or documents and converts them into a structured form. It is used to automatically extract entities, relations, and attribute values and turn them into tables or databases.
In a materials paper, composition, process conditions, and property values are automatically extracted to build a dataset.
Machine Learning Potential · ML Potential
A model trained to rapidly predict physical quantities such as energy, forces, and stress from atomic arrangements. It is used for molecular dynamics or structure exploration at a much lower cost than quantum-mechanical calculations.
Large-scale structure exploration is performed with a Machine Learning Potential · ML Potential to find candidate phase equilibria under high pressure.
Virtual Reference Electrode
A virtual sensor concept that estimates an internal potential or state from measurable cell signals without inserting an additional real electrode. It is used to reconstruct electrode potentials that are difficult to observe by applying machine learning.
In battery degradation diagnosis, the Virtual Reference Electrode is used to estimate the anode potential.