Active Learning (Machine Learning)
A training strategy where the model identifies which unlabeled examples would be most informative to label next, reducing the amount of labeled data needed to reach good performance.
Glossary
Plain-language definitions for AI terms, research methodology jargon, and model names used across the site.
A
AI Agent
An AI system that takes sequences of actions autonomously — calling tools, browsing the web, writing and running code — to complete a multi-step goal without constant human oversight at each step.
Attention Mechanism
The component in transformer models that lets the model weigh which parts of the input are most relevant to each word it generates — the core innovation behind modern LLMs.
B
Bayesian Inference
A statistical framework that updates prior beliefs with observed data to produce posterior probabilities — contrasted with frequentist statistics, and increasingly used in AI model training, adaptive trials, and uncertainty quantification.
Benchmark
A standardized test or dataset used to measure and compare AI model performance — when a paper claims 'state-of-the-art results,' it usually means beating prior models on a specific benchmark.
Bias (AI/ML)
Systematic errors in AI model outputs that arise from skewed training data, flawed problem framing, or optimization objectives that don't match real-world goals — distinct from statistical bias and a major concern when deploying AI in scientific and clinical settings.
C
Chain-of-Thought Prompting
A prompting technique that asks an AI model to show its reasoning step by step before giving an answer, which improves accuracy on complex tasks — especially math, logic, and multi-step analysis.
Citation Graph
A network of academic papers connected by citation links — the data structure underlying tools like Connected Papers, Litmaps, and ResearchRabbit that visualize how papers relate to each other.
Confidence Interval
A range of values consistent with the observed data at a given confidence level — more informative than a p-value because it shows both the direction and plausible magnitude of an effect.
Confounding
A variable that causally affects both the exposure and outcome in a study, creating a spurious association — the central challenge in observational research and a key source of unreliable findings that AI tools can amplify rather than correct.
Context Window
The maximum amount of text an LLM can process in a single interaction — everything you send it plus everything it has generated so far must fit within this limit.
Crystal Structure Prediction
Computationally determining the stable three-dimensional atomic arrangement of a crystalline material from its chemical composition alone — a key challenge in materials discovery and drug development.
D
Data Augmentation
Techniques that artificially expand a training dataset by applying label-preserving transformations to existing examples — random rotations and flips for images, paraphrase generation for text, noise injection for signals — reducing overfitting without collecting new data.
De novo design
Generating entirely new molecular structures from scratch to meet a set of target properties — rather than selecting from existing compound libraries.
Density functional theory (DFT)
A quantum-mechanical computational method for calculating electronic structure and properties of atoms, molecules, and materials — the dominant method for generating training data for AI materials and chemistry models.
Diffusion Model
A class of generative AI model that learns to create new data (images, proteins, molecules, weather states) by learning to reverse a gradual noising process — the basis of tools like Stable Diffusion, RFdiffusion, and GenCast.
E
Effect Size
A standardized measure of how large a difference or relationship is, independent of sample size — the quantity that tells you whether a statistically significant result is actually meaningful.
Embedding
A numerical vector representation of an object (a word, protein, molecule, or document) that captures its meaning or properties in a form AI models can work with — objects that are semantically similar tend to have similar embeddings.
Ensemble Forecasting
Running many slightly different versions of a forecast model to produce a range of possible outcomes rather than a single prediction — the standard approach for quantifying forecast uncertainty in meteorology.
F
Federated Learning
A machine learning approach where a model is trained across multiple data-holding institutions without the raw data ever leaving each site — particularly important for medical research where patient data cannot be shared.
Fine-Tuning
Adapting a pre-trained AI model to a specific task or domain by continuing to train it on a smaller, task-specific dataset — rather than training a new model from scratch.
Formation energy
The energy change when a compound is formed from its constituent elements in their standard states — a key indicator of thermodynamic stability and the primary prediction target for AI materials discovery models.
Foundation Model
A large AI model trained on broad data at scale that can be adapted to a wide range of downstream tasks — GPT-4, Claude, Gemini, and AlphaFold 3 are all foundation models.
G
Graph Neural Network (GNN)
A type of neural network designed to operate on graph-structured data — where entities (nodes) are connected by relationships (edges) — used in molecular modeling, weather forecasting, and materials discovery.
Gravitational wave
Ripples in spacetime produced by accelerating massive objects — detected by laser interferometers and increasingly analyzed with AI to extract source parameters from noisy signals.
GWAS (Genome-Wide Association Study)
A statistical approach that scans hundreds of thousands of genetic variants across thousands of genomes to identify variants associated with a disease or trait — the foundation of modern human genomics and polygenic risk research.
H
Hallucination (AI)
When an AI model generates text that sounds confident and plausible but is factually incorrect — a fundamental limitation of large language models that matters enormously in research contexts.
I
In-Context Learning
An LLM's ability to learn a new task from examples provided directly in the prompt, without any weight updates — the mechanism behind few-shot and zero-shot prompting.
Interatomic Potential
A mathematical function that approximates the energy of a system of atoms given their positions — used in molecular dynamics and materials simulation to avoid the cost of full quantum mechanical calculations.
K
Knowledge Graph
A structured representation of entities and the relationships between them — used in biomedical databases, literature mining tools, and AI systems that need to reason over facts rather than text.
L
Large Language Model (LLM)
An AI model trained on vast text corpora to predict and generate language — the technology behind ChatGPT, Claude, and Gemini, increasingly used in research for writing, summarization, and literature synthesis.
Latent Space
The compressed, continuous representation space a generative AI model uses internally — where similar molecules, proteins, or data points cluster together, and where interpolation and search can discover new candidates.
M
Meta-Analysis
A statistical technique that pools quantitative results from multiple independent studies to produce a combined effect estimate with higher statistical power than any individual study.
Molecular Dynamics (MD)
A computational simulation method that models how atoms and molecules move over time by numerically integrating Newton's equations of motion — used extensively in chemistry, biochemistry, and materials science.
Molecular fingerprint
A fixed-length binary or count vector that encodes which structural features are present in a molecule — the standard way to convert chemical structures into numerical inputs for machine learning models.
Multi-omics
The integration of data from multiple 'omic' layers — genomics, transcriptomics, proteomics, metabolomics — to build a more complete picture of biological state than any single layer provides.
Multimodal Model
An AI model that processes and generates multiple data types — text, images, audio, or structured data — within a single model, rather than using separate specialized models for each modality.
N
Named entity recognition (NER)
An NLP task that identifies and classifies named things in text — people, organizations, places, dates, gene names, drug names — enabling structured extraction from unstructured documents.
Natural Language Processing (NLP)
The field of AI focused on enabling computers to understand, interpret, and generate human language — the foundation for LLMs, semantic search, text extraction, and most AI research tools.
Numerical Weather Prediction (NWP)
The traditional physics-based approach to weather forecasting, which solves mathematical equations representing atmospheric dynamics on a grid — the method AI weather models are increasingly being benchmarked against.
O
Open Weights
AI models whose trained parameters are publicly released for download and local deployment — enabling privacy, customization, and offline use — but whose training data, code, and full methodology may not be disclosed.
Overfitting
When an AI model learns the training data too specifically and fails to generalize to new data — a fundamental challenge when training models on small scientific datasets.
P
p-value
The probability of observing results at least as extreme as the data if the null hypothesis were true — widely misunderstood, widely misused, and insufficient on its own for evaluating scientific evidence.
Perplexity (Language Model Metric)
A measure of how well a language model predicts a text sample — lower perplexity means the model assigned higher probability to the actual words, indicating better predictive performance.
Pre-Registration
The practice of publicly registering a study's hypotheses, methods, and analysis plan before data collection, to prevent HARKing (Hypothesizing After Results are Known) and selective reporting.
Preprint
A scientific manuscript posted publicly before formal peer review — enables rapid dissemination of research findings and is standard practice in physics, biology, and economics, though it creates challenges for AI tools that cannot distinguish peer-reviewed results from unreviewed claims.
PRISMA
Preferred Reporting Items for Systematic reviews and Meta-Analyses — the standard reporting checklist and flow diagram that journals require for systematic review publications.
Prompt Engineering
The practice of designing inputs to an AI language model to reliably get better, more accurate, or more useful outputs — a practical skill for researchers using LLMs in their work.
Protein Folding
The process by which a protein's amino acid sequence determines its three-dimensional structure — a 50-year unsolved problem that AlphaFold largely solved in 2020 using deep learning.
Protein Language Model
An AI model trained on large databases of protein sequences — treating amino acids like words — to learn representations that capture evolutionary and structural information useful for property prediction and design.
Proteomics
The large-scale study of the full complement of proteins expressed by a cell, tissue, or organism — connecting genomic information to biological function, and a field where AI is accelerating both structure prediction and abundance analysis.
Q
Qualitative data analysis (QDA)
The systematic interpretation of non-numerical data — interviews, field notes, documents, images — to identify patterns, themes, and meanings. AI tools can assist with coding and thematic analysis but require careful human oversight.
R
Reanalysis Data
A consistent historical record of atmospheric conditions produced by running a modern weather model over decades of past observations — the primary training data for AI weather models like GraphCast.
Reinforcement Learning from Human Feedback (RLHF)
A training technique that uses human preference judgments to fine-tune AI models toward being more helpful, harmless, and honest — the key method behind how ChatGPT, Claude, and similar assistants behave.
Reproducibility
The ability of independent researchers to obtain the same results using the same data and methods — distinct from replicability (same results with new data) and a growing concern in AI research where model and training variability make reproduction difficult.
Retrieval-Augmented Generation (RAG)
An approach that grounds an LLM's responses in retrieved documents — the model searches a document set first, then generates its answer based on what it actually found, reducing hallucination.
Retrosynthesis
A planning strategy in organic chemistry that works backward from a target molecule to identify feasible synthetic routes — AI tools now automate the search for these routes using reaction databases.
S
Semantic Search
Search that retrieves results based on meaning rather than exact keyword matching — how tools like Elicit and Semantic Scholar find relevant papers even when they use different terminology than your query.
Simulation-based inference
A class of machine learning methods that infer the parameters of a scientific model by learning from simulations rather than requiring a tractable likelihood function — particularly useful in physics and cosmology.
SMILES notation
A compact text format for encoding molecular structures as strings — widely used to represent chemical compounds in AI training data and tool inputs.
Statistical Power
The probability that a study will detect a true effect if one exists — determined by sample size, effect size, and alpha level. Underpowered studies waste resources and produce unreliable results.
Structure-based drug design
Designing drug molecules by exploiting the 3D structure of a biological target — using shape and chemistry of the binding site to guide which compounds to synthesize or screen.
Structured Data Extraction
The process of systematically pulling specific, predefined data fields from a set of documents — in research contexts, used to populate tables comparing outcomes, methods, or sample characteristics across papers.
Survey data (astronomy)
Large-scale astronomical observations systematically covering large areas of sky — the primary data source for many AI astronomy applications, producing catalogs of billions of objects.
Symbolic regression
A machine learning technique that searches for mathematical equations that fit a dataset — producing interpretable formulas rather than black-box predictions.
Synthetic Data
Artificially generated data that mimics the statistical properties of real data — used to augment small training datasets, preserve privacy in sensitive domains, and test models on scenarios that don't yet exist in real data.
Systematic Review
A structured synthesis of all available evidence on a specific research question, following a pre-registered protocol to minimize bias — the highest-quality evidence type in evidence-based medicine and policy.
T
Tokenization
The process of splitting text into tokens — the basic units an LLM processes — which are roughly word pieces, not whole words, explaining why a 10,000-word document uses more than 10,000 tokens.
Topic Modeling
An unsupervised machine learning method that discovers recurring themes in a text corpus — useful for exploring large collections of survey responses, interview transcripts, or documents.
Transfer Learning
Using knowledge a model learned from one task or dataset to improve performance on a different but related task — the principle behind why pre-trained models can be adapted with far less data than training from scratch.
Transformer
The neural network architecture behind most modern AI — including GPT-4, AlphaFold, and ESM-2 — built around an attention mechanism that lets the model weigh relationships between all parts of its input simultaneously.
V
Vector Database
A database that stores and searches over embedding vectors — the data infrastructure underlying retrieval-augmented generation and semantic search in AI research tools.
Virtual screening
Computationally ranking large libraries of compounds against a biological target to identify candidates worth testing experimentally — shortlisting millions of molecules before spending resources in the lab.
Z
Zero-shot classification
Classifying text into categories without training examples — you provide label names in plain language, and the model applies them based on semantic understanding alone.