Materials & AI · glossary
Met an unfamiliar term while reading? Start with the concept.
37 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.
Linear Regression · 최소제곱 · OLS
A model that predicts the output as a linear combination of the inputs, minimising the sum of squared residuals.
Approximating the lattice constant as a first-order expression in the composition ratios.
Random Forest · RF
An ensemble model that averages or votes over decision trees trained to differ from one another.
Predicting hardness from a few hundred alloy measurements.
Gradient Boosting · XGBoost · LightGBM
An ensemble method that adds trees which successively correct the residuals of the models before them.
Predicting yield from tabular process data.
Gaussian Process · GP · 크리깅 · Kriging
A model that places a probability distribution over functions themselves, so it returns a prediction together with its uncertainty.
Reporting large predictive uncertainty in regions where few experiments exist.
Multilayer Perceptron · MLP · 신경망
The basic neural network, stacking linear transformations and non-linear activations in layers to learn complex relationships.
Learning the non-linear relationship between several process variables and a property.
Graph Neural Network · GNN
A neural network that learns from relationships in data represented as nodes and connections.
Predicting properties by treating atoms as nodes and neighbouring pairs as connections.
Transformer · Attention · 자기어텐션
A model built on attention, which learns how much each input element should refer to the others.
Used as the backbone for models over composition sequences or over sentences in papers.
PCA · Principal Component Analysis
A linear dimensionality reduction that finds the directions of largest variance and re-expresses the data on a few axes.
Reducing dozens of descriptors to two axes to look at how compositions are distributed.
Clustering · k-means · 계층 군집
An unsupervised method that groups similar data together.
Grouping measured spectra by type to separate families of samples.
t-SNE · UMAP · Manifold Learning
Visualising high-dimensional data in low dimensions while preserving local neighbourhoods. Distances and sizes on the resulting plot need care in interpretation.
Drawing crystal-structure embeddings as a two-dimensional map to inspect the clusters.
Autoencoder · AE
A neural network trained to compress an input to a low dimension and reconstruct it, yielding a representation.
Compressing spectra and using the latent representation to find unusual samples.
Variational Autoencoder · VAE
An autoencoder whose latent space is learned as a probability distribution, so new samples can be generated from it.
Moving through the latent space to produce new candidate compositions.
Diffusion Model · DDPM
A model that learns to reverse a process of gradually adding noise, so it can generate samples out of noise.
Starting from noise and generating candidate crystal structures.
Physics-Informed Neural Network · PINN
A neural network trained with the governing equations or boundary conditions inside its loss, so it respects the physics.
Learning a concentration profile that satisfies the diffusion equation.
Symbolic Regression
Searching over expressions themselves to find an interpretable relationship that fits the data.
Finding a relationship between dimensionless numbers from measurements.
Genetic Algorithm · GA · 진화 알고리즘
An optimisation method that repeats selection, crossover and mutation to move towards good solutions.
Evolving layer counts and thicknesses to reach a target reflectance.
Reinforcement Learning · RL
Learning a good policy by repeatedly experiencing the reward that follows an action.
Learning the order and the conditions of a synthesis process against a reward.
Multi-objective Optimization · Pareto Front · 파레토 최적
Optimisation that considers several conflicting objectives at once and finds the set of solutions that cannot be improved without making something else worse.
Selecting a set of candidates that weighs strength and ductility together.
Diffractive Neural Network · D2NN
A neural network that physically implements neural-network operations by using diffraction and interference of light. The propagation process itself acts like a computational layer, and the parameters of the optical structure are optimized through training.
In an optical classifier using metasurfaces, a diffractive neural network directly processes the input image.
Equivariant Graph Neural Network
A graph neural network whose output changes according to the corresponding transformation law when a specific symmetry transformation is applied to the input structure. It is suitable for predicting molecular and crystal structures by reflecting physical symmetries such as rotation and translation.
An equivariant graph neural network is used when predicting charge density or forces while preserving the rotational symmetry of a crystal structure.
Contrastive Learning
A method that learns representations by bringing similar samples closer together and pushing different samples farther apart. It is often used to build embeddings that capture structural relationships well even when labels are scarce or unavailable.
It learns material representations by aligning crystal structures and text descriptions close to each other in the same space.
Machine Learning Regression
A supervised learning method that learns a function between input variables and target values in order to predict continuous values. Unlike classification, its goal is to match the magnitude of the value itself.
It is used when predicting continuous properties such as band gap from composition or structural features.
Domain Adaptation
A technique that adapts a model trained in one environment to a target domain with a different distribution or conditions. The key is to reduce the distribution gap between the source domain and the target domain.
It is used when a defect detection model trained in laboratory A is calibrated to field data from a different process condition.
Joint Generation
A method in which two or more interrelated representations or target variables are created simultaneously in a single generation process. It can reflect the structural correlations among the variables together.
Crystal structures and local electronic descriptors can be jointly generated to design candidate materials.
Line Graph Neural Network
A graph neural network that learns by converting edge relationships in a graph back into nodes. It can express relational information more directly in materials-property problems where linear connectivity structures such as atom-bond-angle are important.
A Line Graph Neural Network can be used to predict electronic structure by reflecting bond-angle information in a crystal structure.
LoRA · Low-Rank Adaptation
LoRA · Low-Rank Adaptation is a fine-tuning technique that learns only a small number of parameters by adding low-rank matrices instead of directly changing the weights of a large model significantly. It can perform domain adaptation while reducing computational cost and memory usage.
When adapting an existing MLIP with a small amount of materials data, LoRA can be used to lower the training burden.
Multi-head Fine-tuning
Multi-head Fine-tuning is a method that places multiple output heads on top of a single shared backbone and fine-tunes the model to learn different prediction tasks or material properties simultaneously. It can leverage both shared knowledge across tasks and specialized predictions.
Multi-head Fine-tuning can be applied so that phonons and thermal and elastic properties are predicted together in the same ML potential.
Recursive Feature Elimination · RFE
This is a feature selection method that repeatedly removes features with low contribution to model performance, leaving only the important variables. It is often used to improve both training performance and interpretability.
When there are many candidate composition descriptors, Recursive Feature Elimination can be used to retain only the key variables.
Combinatorial Optimization
A field that deals with problems of optimizing an objective function among combinations of finite or discrete choices. Because the number of candidates can be very large, approximate solutions or heuristics are often used instead of exact search.
A problem of finding the most stable combination among candidate structures can be viewed as Combinatorial Optimization.
Heuristic
An empirical rule or search strategy used to solve a complex problem quickly. It does not guarantee an optimal solution, but it is useful for reducing computational cost and efficiently finding a good solution.
When exploring a broad composition space, a Heuristic is used to select promising candidates first.
Conditioning
A method that adds auxiliary information to a model input and adjusts the prediction or generation process according to that information. It is useful when reflecting contextual information such as metadata, environmental variables, or system characteristics.
By conditioning a catalytic atomistic potential on system metadata, different reaction systems can be distinguished.
Support Vector Regression · SVR
Support Vector Regression · SVR is a method that extends support vector machines to regression problems, controlling the complexity of the prediction function while allowing errors within a certain range. Nonlinear relationships can be expressed through kernel functions.
In spectroscopic property prediction, support vector regression can be used to model nonlinear relationships in small datasets.
Flow Matching
A generative modeling method that directly learns a continuous vector field or flow toward the data distribution. It focuses on matching the transformation path itself rather than on stochastic noise reversal.
In generative crystal structure prediction, it appears when candidate structures are generated by learning a structural vector field.
Equivariance
It is the property that, when a transformation is applied to an input, the output changes in a corresponding way as well. It is an important concept when reflecting symmetry in a model.
An equivariance model is used in which, when a molecular structure is rotated, the predicted force vector also rotates in the same way.
Singular Value Decomposition · SVD
A linear algebra technique that decomposes an arbitrary matrix into the product of orthogonal matrices and a diagonal matrix. It is often used to separate the main structure of data into low-dimensional components or to construct matrix approximations.
The weight matrix of an LLM can be decomposed by singular value decomposition to perform a low-rank approximation and fine-tuning initialization.
Latent Space
Latent space is a space in which high-dimensional data are compressed and represented in a lower-dimensional continuous space. It is used in generative models and embeddings to handle the hidden structure of data.
Exploring catalyst candidates in latent space can continuously suggest new composition combinations.
Orthogonal Transformation
An orthogonal transformation is a linear transformation that preserves vector length and angle. It includes rotations and reflections, and is used to change representation while maintaining geometric structure in numerical settings.
Orthogonal transformation can be used when transforming candidate material representations while preserving the structure of latent space.