Feature and Descriptor
Feature · Descriptor · 피처
A measured value or a numerical representation of an object, used as model input.
In materials terms
scikit-learn 용어집 ↗Using the mean atomic radius as a descriptor of a composition.
Materials & AI · glossary
Materials and AI glossary
Met an unfamiliar term while reading? Start with the concept.
20 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.
Regression
Regression
A supervised problem whose target is a continuous numerical value.
In materials terms
scikit-learn 용어집 ↗Predicting the band gap of a material in eV.
Classification
Classification
A supervised problem that predicts which category an input belongs to.
In materials terms
scikit-learn 용어집 ↗Separating sample images into sound and defective.
Overfitting
Overfitting
The state where a model fits its training data well but performs worse on new data.
In materials terms
scikit-learn 용어집 ↗Matching known alloys closely while the error grows for new compositions.
Cross-validation
Cross-validation · CV
A validation method that splits the data several times and repeats training and evaluation.
In materials terms
scikit-learn 교차검증 문서 ↗Keeping related samples in the same group so no evaluation information leaks.
Hyperparameter
Hyperparameter
A model or training setting chosen outside the learning process itself.
In materials terms
scikit-learn 용어집 ↗Choosing tree depth using validation data.
Bias-Variance Tradeoff
Bias-Variance Tradeoff · 편향-분산 절충
The trade-off between error from a model being too simple and error from it being too sensitive to the data.
In materials terms
scikit-learn 용어집 ↗Raising the polynomial degree lowers the bias but raises the variance.
Learning Curve
Learning Curve
A curve showing how performance changes with the amount of data or the progress of training.
In materials terms
scikit-learn 교차검증 문서 ↗Judging whether collecting more data would help performance.
Regularization
Regularization · Ridge · Lasso · 정규화
Adding a term that limits the size of the coefficients in order to hold back overfitting.
In materials terms
scikit-learn 선형 모델 문서 ↗Keeping only the descriptors that matter with Lasso when there are many of them.
Third-party Verification
Third-party Verification
It is the confirmation of results and procedures by a third party rather than the party that develops or operates the system. It is used to create a reliable evaluation framework by increasing independence and transparency.
In materials terms
Anthropic update on independent evaluators \ Anthropic ↗The prediction performance of a materials AI model can be re-verified by an external institution to increase reliability.
Employee-like Access
Employee-like Access
A method of granting external evaluators or collaborators system access privileges similar to those of internal employees. It enables deeper inspection and reproducible analysis, but security and the scope of control must be designed together.
In materials terms
Anthropic update on independent evaluators \ Anthropic ↗An external validation team performs more detailed model safety testing with Employee-like Access.
Auditability
Auditability
The property that the input, output, and decision process of a system can be traced and reviewed later. It is important for accountability, reproducibility, and security verification.
In materials terms
Developing Gemini 4 Argon responsibly and safely ↗In an autonomous laboratory, auditability is ensured by recording which model recommended the experimental conditions and on what basis.
Independent Evaluation
Independent Evaluation
A method in which a third party separate from the developer evaluates performance, safety, and quality using the same criteria. It is important for increasing the objectivity and reliability of the evaluation.
In materials terms
Anthropic update on independent evaluators \ Anthropic ↗A new materials recommendation model can reduce overestimation through Independent Evaluation.
Catastrophic Forgetting
Catastrophic Forgetting
A phenomenon in which previously learned knowledge is rapidly lost when training on a new task or new data. It is often a problem in transfer learning and continual learning.
In materials terms
AQCat25: unlocking spin-aware, high-fidelity machine learning potentials for heterogeneous catalysis | npj Computational Materials ↗When fine-tuning an MLIP for catalysis with new composition data, the accuracy of the previous catalyst family can decrease.
Red Teaming
Red Teaming
A validation method that deliberately seeks out weaknesses, bypass possibilities, and failure modes in a model or system. It is used to check safety, robustness, and misuse potential.
In materials terms
Partnering with Accenture on embedded evaluation \ Anthropic ↗Extremely adversarial inputs are given to an autonomous experiment recommendation model to induce incorrect condition suggestions.
Self-supervised Learning
Self-supervised Learning
A method that learns representations by creating training signals from the input data itself without ground-truth labels. It can use large amounts of unlabeled data to first learn general features.
In materials terms
WeedNet: a foundation model-based global-to-local AI approach for weed species identification and classification | Nature Communications ↗When labels are scarce in materials microscopy images, Self-supervised Learning is first used to learn representations.
Verification
Verification
The process of checking whether the output of a model or procedure satisfies the intended requirements, constraints, or ground-truth criteria. Unlike evaluation of predictive performance, it examines whether it operates correctly under specified conditions.
In materials terms
[2609.27621] SHRAV: State-Hypothesis-Reason-Action-Verify Framework for Physical Modeling and Inverse Design ↗The reverse-design result is checked in the verification stage to confirm that it satisfies the process constraints.
Benchmark Design
Benchmark Design
It is the work of constructing evaluation datasets and metrics so that they well reflect a specific task and usage context. Its purpose is to reveal strengths and weaknesses under real operating conditions rather than relying on simple average performance.
In materials terms
OMC-bench and AtomBit-OMC: task-aligned benchmarks and robust, interpretable machine-learned interatomic potentials for organic molecular crystals | npj Computa ↗Separate benchmarks are designed for each task, such as structural relaxation and polymorph ranking of organic molecular crystals.
Unsupervised Deep Learning
Unsupervised Deep Learning
An unsupervised deep learning method learns the structure, patterns, and latent representations of data without ground-truth labels. It is often used for reconstruction, clustering, and representation learning.
In materials terms
Unsupervised deep learning for limited-angle STEM-EDX tomography: application to 3D chemical analysis of phase-change memory devices ↗Unsupervised deep learning can be applied to reconstruct missing information in limited-angle tomography.
Closed-loop Optimization
Closed-loop Optimization
An optimization method that repeatedly explores and improves while immediately reflecting experimental or observational results in the next decision. Model prediction, execution, and feedback form a cyclic structure.
In materials terms
An autonomous lab for data-driven homogeneous catalysis | Nature Communications ↗In an autonomous laboratory, the process of recommending the next synthesis conditions and then updating the model again with the experimental results is Closed-loop Optimization.