Machine Learning
Machine Learning · ML · 기계학습
Learning patterns from data in order to predict or classify values for new inputs.
In materials terms
Predicting the strength of a material from its composition and processing conditions.
Materials & AI · glossary
Materials and AI glossary
Met an unfamiliar term while reading? Start with the concept.
44 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.
Google ML 용어집 ↗
Supervised Learning
Supervised Learning
Learning the relationship between inputs and outcomes from data where both are paired.
In materials terms
Using conductivity measured for each composition as the answer to learn from.
Unsupervised Learning
Unsupervised Learning
Finding similarity or hidden structure in data without any answers to learn from.
Deep Learning
Deep Learning · DL · 심층학습
Using neural networks of several layers to learn a data representation and a prediction rule together.
Large Language Model
Large Language Model · LLM
A model trained on large text corpora to process and generate language. Whether its output is factual has to be checked separately.
In materials terms
Extracting synthesis conditions from a paper and then checking them against the original text.
Generative AI
Generative AI · 생성 모델
AI that produces new candidates — text, structures and so on — from a distribution or pattern it has learned.
In materials terms
Proposing candidate crystal structures that match a target property.
Embedding
Embedding · 벡터 표현
Representing an object’s features or relationships as a vector of numbers.
Inference
Inference · 예측
Applying a trained model to a new input to obtain an output.
In materials terms
Predicting the properties of a new composition with a model that has finished training.
Foundation Model
Foundation Model · 기반 모델
A large model pre-trained broadly on wide data and then adapted to several downstream tasks.
In materials terms
Fine-tuning an atomistic model pre-trained across many element systems onto one specific alloy.
Transfer Learning and Fine-tuning
Transfer Learning · Fine-tuning · 미세조정
Moving a model trained on one task to another task with little data and training it further there.
In materials terms
Adjusting a model trained on large-scale DFT data using a handful of measured values.
Multimodal Learning
Multimodal Learning
A method of training a single model by inputting data of different forms together. It combines complementary information from each modality to improve prediction performance or representational power.
In materials terms
It predicts material properties by feeding both the sequence representation and the 3D structure of a polymer together.
bug bounty
bug bounty
A system that pays rewards to people who discover and report vulnerabilities in software or systems. It is used to encourage external researchers to participate in security testing.
In materials terms
If you find a vulnerability in the research AI platform, you can receive a reward through the bug bounty program.
Interactive Simulation
Interactive Simulation
A simulation in which the simulation state changes immediately according to input from a user or an agent. It is not one-way playback, but explores the system by exchanging feedback.
In materials terms
It can be used in a materials design interface where the results are updated immediately when experimental conditions are changed.
World Simulation
World Simulation
A simulation that reproduces the dynamics and interactions of the real world or a virtual environment. By including physical responses, time evolution, and agent behavior together, it makes it possible to explore a system experimentally.
In materials terms
In a materials process, a virtual simulator can be constructed that reflects temperature changes and equipment operation together.
Model Spec
Model Spec
A normative document that organizes the behavioral principles, priorities, and prohibitions that an AI model must follow. It defines not only the quality of the model's responses but also safety, honesty, and human controllability.
In materials terms
The response policy of an autonomous laboratory agent is designed in accordance with the Model Spec.
World Model
World Model
A generative model that predicts both changes in an environment’s state and observations, while also simulating the results of actions within it. It helps an agent internally construct a “virtual world” and reason before directly exploring the real world.
In materials terms
It can be used to virtually simulate how the microstructure will change when material processing conditions are altered.
Vision Reasoning
Vision Reasoning
The ability to interpret information contained in images or videos and, based on that content, carry out logical judgments or subsequent tasks. It includes understanding visual context and relationships beyond simple object recognition.
In materials terms
It is used when interpreting defect distribution in a microscope image to determine the next analysis step.
Evidential Deep Learning
Evidential Deep Learning
A method that learns not only the model’s predicted value but also the strength of the evidence for that prediction, in order to estimate uncertainty. By directly predicting distribution parameters, it can express confidence when data are scarce or ambiguous.
In materials terms
When materials data are limited, evidential deep learning can provide both the predicted value and the uncertainty for candidate compositions.
Network Isolation
Network Isolation
A security method that separates a system or training environment from an external network to restrict access. It is used to reduce the risk of data leakage, external intrusion, and unauthorized transmission.
In materials terms
When training a large model on sensitive materials data, the training server is separated from the external internet.
User Interface Generation · UI generation
User Interface Generation · UI generation
A technology that automatically creates screen layouts, widget placement, and interactive elements based on natural-language requirements or task context. It includes generating output in a directly usable interface form rather than a text response.
In materials terms
A dashboard screen draft can be automatically generated from a request to view experimental data.
Access Control
Access Control
A principle that restricts and manages whether users or systems can access certain resources, data, or functions according to permissions. Least privilege and role-based authorization are representative approaches.
In materials terms
Sensitive experimental data and automation equipment are made available only to authorized researchers through access control.
Governance
Governance
An AI management framework that sets the scope of use, responsibility, control, and audit procedures for an AI system. It includes not only technical performance but also policy compliance and risk management.
In materials terms
When AI is introduced into an autonomous laboratory, governance procedures must be designed together with model accuracy.
Jailbreak
Jailbreak
An act of bypassing the safety constraints or policies of a conversational AI to induce outputs that are originally not allowed. It appears in various forms such as prompt manipulation, role-play, and induced instruction conflicts.
In materials terms
In a materials literature summarization tool, a jailbreak attempt may be used to elicit prohibited experimental procedures.
Prompt Injection
Prompt Injection
An attack method in which malicious or unintended instructions are inserted into the directives that a conversational AI or agent follows, inducing behavior that departs from the original intent. It occurs when the model fails to properly distinguish between input and instruction.
In materials terms
A literature summarization agent must be protected so that it does not fall for hidden instructions on a webpage and draw incorrect conclusions.
Tool Use
Tool Use
A model’s way of performing tasks by calling tools such as external calculators, search engines, software, or equipment control APIs. It supplements calculations or execution that are difficult to resolve through the model’s internal reasoning alone.
In materials terms
In crystal structure analysis, it calls a diffraction pattern processing program to find peaks.
Agentic AI
Agentic AI
An AI system that combines tool use, planning, and reasoning to carry out goals with relatively high autonomy. It goes beyond simple question answering and can interact with its environment while selecting the next action on its own.
In materials terms
It sequentially calls beamline control software and measurement tools to perform the initial experimental procedure.
Scientific Reasoning
Scientific Reasoning
A thinking process that connects observation, hypotheses, and evidence to derive explanations or predictions. It is important for interpreting cause and mechanism consistently beyond data patterns.
In materials terms
It explains the cause of property changes through the relationship between structure and composition.
Vulnerability Scanning
Vulnerability Scanning
A method of inspection that automatically finds known security vulnerabilities in software or systems. It checks package dependencies, libraries, configurations, and code patterns to identify risk factors early.
In materials terms
The open-source dependencies of the materials analysis pipeline are checked with Vulnerability Scanning before deployment.
Small Language Model
Small Language Model
A language model with a relatively small number of parameters and a low computational load, with fast response and low operating cost as advantages. Although its capabilities may be more limited than those of a large model, it is highly practical for certain tasks.
In materials terms
A Small Language Model can be attached to a materials information workflow that repeatedly performs large-scale literature searches.
User Interface · UI
User Interface · UI
A contact point such as a screen, menu, or input method through which a user interacts with software or a system. It determines not only functional power but also ease of use and workflow.
In materials terms
A researcher uses the interface of an AI tool to more easily carry out literature searches and data organization.
Joint-Embedding Predictive Architecture · JEPA
Joint-Embedding Predictive Architecture · JEPA
This is a self-supervised learning architecture that is trained to predict part of the input in a latent representation space. It focuses on learning meaningful representations rather than directly reconstructing the original.
In materials terms
Joint-Embedding Predictive Architecture can be applied to pretraining that predicts the masked parts of 3D atomic structures.
Deployment
Deployment
The process of placing a trained model or system into an operating environment so it can be used by actual users, services, or infrastructure. It includes not only performance, but also access approach, scalability, security, and maintainability.
In materials terms
When deploying a model to a laboratory automation system, the stability of the connection with the equipment must also be checked.
Brand Safety
Brand Safety
A concept that ensures advertisements are not exposed in contexts that are inappropriate or harmful to the brand. It manages content context, level of expression, and topic suitability together.
In materials terms
In a material recommendation system, it checks so that an advertisement for a specific product is not attached to a sensitive conversation context.
Scale Invariance
Scale Invariance
A property in which the core relationship changes little even when the input size, unit, or resolution changes. It is important when a model or physical law must operate consistently under scale changes.
In materials terms
It is important when designing the system so that phase-field prediction performance is maintained even when the grid size differs.
Physics-Informed Neural Operator · PINO
Physics-Informed Neural Operator · PINO
A neural operator trained to satisfy physical laws. It can rapidly approximate operators that output an entire function or field while the input conditions change.
In materials terms
It can be used to quickly predict the spatial distribution of a phase-field solution.
Agent
Agent
A software or system that observes the environment and selects actions to perform tasks based on a given goal. It can automate complex work through tool calls and stepwise task execution.
In materials terms
A research assistant tool that performs literature searches, organizes experimental records, and enters data can be regarded as an Agent.
Safety Evaluation
Safety Evaluation
A process of evaluating whether a model or system produces unintended harmful outputs, policy violations, or reliability problems. It checks risk and failure modes together with functional performance.
In materials terms
When introducing an LLM for summarizing materials data, it is used to check in advance for sensitive information leakage and incorrect citations.
Context Window
Context Window
The maximum length of tokens that a model can reference in a single input. A longer window preserves longer context for reasoning, but it also increases computational cost and the possibility of error accumulation.
In materials terms
It is important when comparing materials synthesis conditions by putting several long papers and experimental records into the model at once.
Hypothesis Generation
Hypothesis Generation
The process of proposing a new, testable explanation or candidate idea based on observed data or the literature. It creates the starting point for the next experiment or calculation in exploratory research.
In materials terms
AI generated a hypothesis for a new composition candidate based on the literature and experimental data.
Misalignment
Misalignment
A state in which a model's actual behavior diverges from the goals, rules, or safety requirements intended by the developer or user. Separate from fluency of output, it can appear as hidden goal pursuit, distorted instruction following, or evasion.
In materials terms
A classification model is checked when it sounds plausible in its explanation but in fact misleads the user by bypassing prohibited features.
Secure Enclave · Trusted Execution Environment
Secure Enclave · Trusted Execution Environment
An environment that protects and executes code and data inside a processor or an isolated hardware region. It increases confidentiality by making it difficult for external systems to directly read sensitive information in memory or during computation.
In materials terms
AI inference containing sensitive research data is performed in a Secure Enclave · Trusted Execution Environment to reduce access risk.
Unsupervised-supervised learning
Unsupervised-supervised learning
Unsupervised-supervised learning is an approach that first finds the structure or clusters of data through unsupervised learning and then uses the results to perform supervised learning for classification or prediction. It is useful for jointly carrying out exploration and discrimination when labels are scarce.
In materials terms
One can first cluster candidate regions of a phase transition and then use a classifier to distinguish Majorana states.
Interpretable Machine Learning
Interpretable Machine Learning
A machine learning approach that makes it possible for people to understand not only a model's predictions but also the reasons for them and how it works. It uses methods such as feature importance, rules, and partial dependence together.
In materials terms
In catalyst activity prediction, Interpretable Machine Learning is used to explain which composition variables and process conditions are important.
In-context Learning
In-context Learning
A method that performs a task using only examples and instructions included in the input context, without updating the model weights. It infers the task rules from the cases provided in the prompt.
In materials terms
In-context Learning can be used to make the system follow the equipment operation sequence by placing example experimental procedures in the prompt.