Starting and learning
Machine learning for research
Researchers use machine learning to build surrogate models for expensive simulations, classify signals and images from experiments, screen materials and molecules, and estimate effects from observational data. Each use needs a clear target, a reproducible split and a fair baseline.
How researchers use machine learning
Predicting structures
Deep learning predicts molecular and protein structures that are slow to determine experimentally. Researchers use the predictions to choose which experiments to run next.
Real exampleAlphaFold (Nature, 2021)DeepMind's paper 'Highly accurate protein structure prediction with AlphaFold' describes a neural network that predicts a protein's 3D structure from its amino acid sequence.
Screening materials
Graph neural networks predict properties such as stability for candidate materials, so only the most promising are simulated in full or synthesised. The model narrows the search, and experiments confirm the result.
Real exampleGNoME (Nature, 2023)Google DeepMind reports that graph networks trained at scale improved the efficiency of discovering stable inorganic crystals by an order of magnitude.
Classifying signals in large datasets
Convolutional networks sort signals in telescope, detector or sensor data into real events and noise, which makes very large surveys manageable. Candidates are then followed up with standard analysis.
Real exampleIdentifying exoplanets with deep learningShallue and Vanderburg trained a convolutional neural network on Kepler telescope signals and identified new planets around Kepler-80 and Kepler-90.
Emulating simulations and forecasts
Models trained on historical data or simulation output produce forecasts far faster than running the full physical model. They are validated against the established method over many cases.
Real exampleGraphCastGraphCast, from Google DeepMind, is trained on reanalysis data and predicts hundreds of weather variables 10 days ahead at 0.25 degree resolution in under a minute.
Checking results for leakage
Machine learning in research needs strict separation of training and test data, or results look better than they are. Reviewers increasingly ask how splits were made and whether the result reproduces.
Real exampleLeakage and the Reproducibility Crisis in ML-based ScienceKapoor and Narayanan found data leakage errors in 17 scientific fields, collectively affecting 329 papers, and propose model info sheets to catch leakage.
Can researchers learn machine learning?
Yes. Researchers usually have the scientific question, the data and statistical training already. Add Python, careful data splitting and the core model families, then apply them to a well-defined question in your field with a fair baseline.
- Python for research dataScripts and notebooks that load, clean and plot your data reproducibly from raw files.Learn it in Python
- NumPy and pandasArrays and tables for measurements, metadata and experimental conditions.Learn it in Data Tools
- Statistics for evaluationUncertainty, resampling and comparing models fairly on held-out data.Learn it in Statistics
- Supervised learning and splitsRegression and classification with group or time-based splits that match how results will be used.Learn it in Machine Learning
- Deep learning for images, signals and graphsNeural networks in PyTorch for data types that tabular models handle poorly.Learn it in Deep Learning
Bringing machine learning into your research
If you already work in the field
Researchers already know their field's questions, data quirks and standards of evidence. Add Python and machine learning practice, then reproduce a published machine learning result from your field before applying the method to your own data.
If you come from data science or ML
Machine learning practitioners joining a research group need the field's vocabulary, how its data is collected and what counts as a meaningful result. Read recent reviews, work closely with domain scientists and agree on evaluation before training anything.
Roles that do this workResearch scientistComputational scientistResearch software engineerPostdoctoral researcher
Tools you will use
| Tool | What you use it for |
|---|---|
| scikit-learn | Provides baselines, cross-validation and metrics for fair, reproducible comparisons. |
| PyTorch | Builds custom neural networks for research models and surrogates. |
| MLflow | Logs parameters, metrics and artefacts for each run so results can be traced and reproduced. |
| DVC | Versions datasets and pipelines alongside code so experiments can be rerun exactly. |
| OpenML | Shares datasets, tasks with fixed splits and results so methods can be compared on the same footing. |
Public data to practise on
- Superconductivty Data: Features derived from the chemical formulas of superconductors, with each material's critical temperature as the target.
- OpenML-CC18 benchmark suite: Curated classification tasks with standard splits for benchmarking a method across many datasets.
Your first project
Download the UCI Superconductivity dataset, predict critical temperature with gradient boosting, and compare it with linear regression using RMSE over five recorded random splits.
Common questions
How is machine learning used in scientific research?
To predict properties and structures, classify signals and images from experiments, emulate expensive simulations and find patterns in large datasets. It usually narrows down what to test, and experiments confirm the result.
How do I make machine learning research reproducible?
Fix and publish your data splits, report baselines, share code and environment details, and check for leakage between training and test data. Model info sheets and data cards help reviewers see what was done.
Do I need a GPU for machine learning research?
Not for tabular models, which run on a laptop. Deep learning on images, sequences or large datasets benefits from a GPU, available through university clusters or cloud notebooks.
Courses for these skills
- Statistics and Probability for Machine LearningBuild probability, inference and experimental-design foundations.41 published lessons · all planned lessons open
- Practical Machine Learning in PythonUnderstand predictive nuisance models and overfitting.51 published lessons · more lessons in production
- Deep Learning with PyTorchStudy neural fitting and loss design.Syllabus available · lessons open as produced
- Maths for Machine LearningConnect derivatives, vectors and physical objectives.26 published lessons · all planned lessons open
New to programming or maths? Start with Python, or follow the machine learning path in order. Published lessons are free to read; compare Free and Pro.