Materials & AI · glossary
Met an unfamiliar term while reading? Start with the concept.
22 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.
Train/Validation/Test Split
Using separate data for fitting the model, for choosing its settings, and for the final performance figure.
Leaving the test data untouched until the result is reported.
Data Leakage · 정보 누출
The problem where information from the evaluation data reaches the training process, making performance look better than it is.
Scaling on the whole dataset first lets information from the test data seep in.
Standardization · Normalization · Scaling
Preprocessing that puts variables on comparable units and ranges to stabilise learning.
Removing the size difference between temperature in degrees and time in hours before fitting.
Class Imbalance
The situation where one category is far more common, so accuracy alone cannot judge performance.
With defective samples at 2 percent, answering “sound” every time still scores 98 percent accuracy.
MAE · RMSE · R²
Measures of the size of the prediction error and of explained variance. They differ in units and in how much outliers move them.
Reading RMSE alongside MAE, because RMSE reacts more strongly to large errors.
Precision · Recall · F1 · ROC AUC · 혼동행렬
Measures that judge a classifier by what it got right and what it missed.
Favouring recall when missing a defect is the costly outcome.
Design of Experiments · DOE · 실험계획법
Arranging conditions in advance so that variable effects and their interactions can be found from few runs.
Using a factorial design to see interactions instead of changing one variable at a time.
Latin Hypercube Sampling · LHS
A sampling scheme that spreads points evenly over each variable’s range, covering the space broadly with few samples.
Exploring a five-variable process space evenly with twenty points.
Extrapolation · Domain of Applicability · 적용 도메인
Prediction in a region the training data does not cover. Error and uncertainty can grow sharply, so the applicable range should be stated alongside.
Treating a prediction for a composition containing an unseen element as indicative only.
FAIR Data · 찾기·접근·상호운용·재사용
Principles for keeping data findable, accessible, interoperable and reusable.
Recording measurement conditions and units as standard fields before publishing.
Materials Project · OQMD · ICSD · AFLOW
Public databases that collect computed and measured materials data with search and download.
Checking the computed properties of candidate compositions before choosing what to make.
Ontology · 개념 체계 · OWL
A formal definition of a field’s concepts and the relationships between them, so that people and programs read the data the same way.
Defining how material, process and property relate so that data from different labs merges onto the same fields.
Materials Ontology · EMMO
An ontology that standardises the materials concepts of substance, structure, process and property, giving data integration and automated interpretation a common reference.
Aligning datasets that recorded the same property under different names onto standard concepts.
Knowledge Graph · RDF · 트리플
Knowledge represented as a graph of nodes and edges holding the relationships between entities, so scattered data can be connected and queried.
Linking composition, process, property and literature to trace the processing route that produced a property.
SPARQL · Graph Query
Querying the triples in a knowledge graph by pattern. Unlike SQL over tables, it asks about relationship paths directly.
Finding “samples containing this element that went through this process” in one path query.
Benchmark · Matbench · 표준 평가셋
An evaluation dataset fixed in advance, with the same data and the same split, so models can be compared.
Comparing your own model against published results on the public split.
Data Cleaning · Data Cleansing
The process of finding and correcting errors, missing values, duplicates, outliers, and other issues in raw data to make it suitable for analysis. It is important for improving data quality and reducing noise before model training.
In automated process data, sensor outliers are removed to improve the performance of a regression model.
Real-time Data Analysis
A method that performs analysis immediately as data are generated, so the results can be quickly reflected in decision-making or control. It reduces latency and enables condition changes and anomaly detection during the experimental process.
In electron beam measurements, real-time data analysis can immediately exclude scans of poor quality.
Cluster Analysis · Clustering Analysis
A statistical method that divides data into multiple groups according to similarity. It is often used to find hidden structures or patterns when there are no prior labels.
Cluster analysis is used on SEM image features to group candidate microstructures that are similar to one another.
Automation
Making a system perform repetitive or rule-based tasks that people used to do. It is used to increase the speed and reproducibility of processes such as experiments, data processing, and decision support.
A robotic system that automatically adjusts synthesis conditions corresponds to nanocrystal synthesis Automation.
Single-cell Profiling
Single-cell profiling is an analysis method that treats individual cells as independent observation units and quantifies cell-specific characteristics such as gene expression, morphology, and image signals. It is useful for revealing heterogeneity and rare cellular states that are hidden by population averages.
In infection experiments, single-cell profiling can be used to compare different responses in each cell.
Segmentation
The process of dividing an image or signal into meaningful regions or objects. It separates boundaries and structures to enable subsequent analysis.
In TEM images, it is used to distinguish particles, defects, and interfaces for automated measurement.