$ aixsci learn
Learn how AI is used in science
Plain-language explainers for the methods in these articles: what each one is, why scientists use it, where it falls short, and where it appears in real research.
Start with the basics
basics
What machine learning is
Machine learning is a way of making computers do useful work by showing them examples and letting them find the patterns, rather than by writing out the rules yourself.
basics
Neural networks
A neural network is a model that learns by adjusting the strengths of connections between many simple units, and it underlies almost all of what is now called artificial intelligence.
basics
Training, testing and using a model
A model is not built by hand but fitted to examples, then checked on examples it has never met, before being set loose on cases nobody knows the answer to.
By field
How AI is used across a whole field, and how each kind of AI is used within it. For named models such as AlphaFold, see models.
field
AI in structural biology and protein science
How artificial intelligence is used in structural biology and protein science: predicting the shapes of proteins, designing new sequences, reading experimental data, and choosing which variants to test next.
field
AI in materials and computational chemistry
How artificial intelligence is used to study and design materials, from predicting what a compound will do to reading microscope images and guiding experiments as they run.
field
AI in astronomy and astrophysics
How artificial intelligence is used to study the sky: sorting the flood of images and spectra that telescopes produce, and reading physical quantities out of them that no one can measure directly.
What the AI was for
The scientific job the AI did, named so that the same work looks the same in every field.
task
Property prediction
Property prediction is the use of a trained computer model to guess a measurable quantity about something that already exists, rather than measuring it in the laboratory or telescope.
task
Classification
Classification is the task of sorting things that already exist into categories that are already known, including the simple question of whether something belongs to one particular group or not.
task
Simulation surrogate
A simulation surrogate is a machine-learning model trained to copy the answers of a slow computer simulation, so that the same question can be asked again and again cheaply.
task
Structure determination
Structure determination is working out the shape and arrangement of an object you cannot see directly, either from indirect measurements of it or from a sequence that describes its parts.
task
Candidate generation
Candidate generation is the step where a computer proposes new things worth trying (recipes, molecules, designs), so that experiments or further calculation can test which ones actually work.
task
Experimental design
Experimental design is the business of deciding which experiment or measurement to do next, rather than running everything or guessing, so that each costly trial is chosen to teach the most.
task
Detection
Detection is the task of finding every instance of a known kind of thing in an image or signal, and saying where each one sits, without being told how many are there.
task
Denoising
Denoising means pulling the signal you care about out of a measurement cluttered with random fluctuation, interference or blur, so that the underlying thing becomes visible.
task
Segmentation
Segmentation is the job of dividing an image or a signal into meaningful pieces, so that a computer can say which part is which object rather than just storing a grid of values.
task
Anomaly detection
Anomaly detection means sifting a mass of routine observations for the few that do not fit, without being told in advance what the odd ones will look like.
task
Literature synthesis
Literature synthesis means pulling claims out of published papers and putting them together, and software can now read thousands of papers and assemble what they report into a single table.
Model families
The kinds of model used, from decades-old statistics to the newest networks.
model
Multilayer perceptron
A multilayer perceptron is the plainest kind of artificial neural network: rows of simple computing units, each connected to every unit in the next row, learning a mapping from inputs to outputs.
model
Convolutional neural network
A convolutional neural network is a learning system built for data laid out on a grid, such as an image, that finds small local patterns and reuses the same pattern detectors everywhere.
model
Transformer
A transformer is a type of neural network that reads a whole sequence at once and learns which parts of it should influence which others, rather than stepping through in order.
model
Random forest
A random forest is a prediction method that grows many simple decision trees on random slices of the data and lets them vote on the answer.
model
Linear model
A linear model is a fitting method that assumes each measured input pushes the answer up or down by a fixed amount, and simply adds those pushes together.
model
Graph neural network
A graph neural network is a model that learns by passing messages between connected things, so what it predicts about any one part depends on its neighbours.
model
Gradient-boosted trees
Gradient-boosted trees are a prediction method that adds up many simple decision trees, each new tree fitted to correct the errors left over by the ones before it.
model
Clustering
Clustering is letting a computer sort data into groups by similarity alone, without being told in advance what the groups are or how many there should be.
model
Protein language model
A protein language model is a program trained on millions of protein sequences, which learns the patterns of how amino acids follow one another much as a text model learns the patterns of words.
model
Gaussian process
A Gaussian process is a way of fitting a smooth curve or surface through scattered measurements that also reports, at every point, how uncertain the fit is.
model
Support vector machine
A support vector machine is a method that learns to draw the widest possible dividing line between groups of examples, and can also be adapted to predict numbers.
model
Recurrent neural network
A recurrent neural network is a learning model that reads a sequence one step at a time and carries a running memory of what it has already seen into the next step.
model
Autoencoder
An autoencoder is a pair of neural networks trained to squeeze data down to a compact summary and then rebuild the original from that summary alone.
model
Probabilistic graphical model
A probabilistic graphical model is a way of writing down which things in a system depend on which others, as a network of links, so that a computer can reason about them all together.
model
Diffusion model
A diffusion model learns to create new data by starting from pure noise and cleaning it up in many small steps, until something that looks like real data remains.
model
Large language model
A large language model is a computer system trained on huge amounts of text that can be given written instructions and will reply in writing, which researchers use as one step inside a larger method.
model
Normalising flow
A normalising flow is a reversible mathematical recipe that bends a simple spread of random numbers into the complicated shape of real data, and can be run backwards again.
model
Generative adversarial network
A pair of learning programs trained against each other, one inventing fake examples and the other trying to spot them, so that the inventor ends up producing convincing imitations of real data.
model
Symbolic regression
Symbolic regression is a way of getting a computer to search for a short formula that fits data, rather than asking a person to guess the formula first.
Ways of learning
How a model is taught: with labelled examples, without them, or by choosing what to learn next.
learning
Supervised
Supervised learning is teaching a computer by example: you show it many inputs alongside the correct answers, and it learns to produce those answers for inputs it has not seen.
learning
Self-supervised
Self-supervised learning trains a model without human-made labels, by hiding part of the data and asking the model to guess what was hidden, so the data effectively labels itself.
learning
Transfer / fine-tuning
Transfer, or fine-tuning, means taking a model that has already learned from a large body of data and adapting it to a new, usually smaller, scientific problem.
learning
Unsupervised
Unsupervised learning is the practice of letting a computer find patterns in data that carries no answers: no labels saying what each item is, only the measurements themselves.
learning
Active learning
Active learning is when a computer model chooses which measurement or experiment to do next, learns from the answer, and then chooses again, instead of being handed a fixed pile of data.
learning
Zero-shot
Zero-shot means using a model on a task straight away, without first showing it any worked examples of that task.
learning
Semi-supervised
Semi-supervised learning trains a model on a small set of examples with known answers together with a much larger set without answers, often by letting the model label the rest itself.
learning
Reinforcement
Reinforcement learning trains a program by letting it act, scoring the results, and nudging it towards whatever earns a better score, so it learns from consequences rather than from worked answers.
Steps in the work
What each step of a piece of research does, whether or not AI is involved.
step
Training
Training is the process of fitting a model to examples, adjusting its internal numbers until its outputs match the answers it is shown, so it can then handle data it has never seen.
step
Inference
Inference is the act of running an already-trained model over fresh data to label, rank or score it, rather than teaching the model anything new.
step
Validation
Validation is the practice of checking what a method produced against something already known to be true, to find out whether its answers can be trusted.
step
Preparation
Preparation is the work of tidying data already collected: cleaning out errors, filtering, putting values on a common scale, attaching labels, so that it can be analysed or used to train a model.
step
Acquisition
Acquisition is the first step in any study: getting hold of the raw data, whether by taking a measurement, running an instrument, downloading a public archive or pulling records out of a database.
step
Interpretation
Interpretation is the work of looking inside a trained model, or at the pattern of its outputs, to learn something about the world rather than just getting a prediction.
step
Representation
Representation is the step where messy raw data is turned into a tidy set of numbers or a network of connected parts, so that a computer program can work with it at all.
step
Simulation
A simulation is a computer model that works out what something would do, step by step, from the rules it is thought to obey, so scientists can watch it happen without doing it.
step
Screening
Screening is sifting a large pile of candidates down to a short list in one pass, by filtering out the unpromising ones or ranking them so the best rise to the top.
step
Optimisation
Optimisation is the practice of searching for the best option by trying something, judging the result, and using that judgement to choose what to try next.
step
Experiment
An experiment is the moment a scientific idea meets the physical world: a sample is actually made, a measurement is actually taken, by human hands or by machine.
step
Generation
Generation is when a computer model makes new candidate objects (a protein sequence, a recipe, a spectrum) that did not exist before, rather than judging ones that already do.
For everyday AI, such as chatbots, writing prompts or using AI at work, see AI for Beginners, our sister site.