Science and engineering
Machine learning for physics
Physicists use machine learning to separate signal from background in particle collisions, to build fast emulators of expensive simulations, and to infer parameters from experiments. Learned models are judged against physical invariants and held-out regimes.
How machine learning is used in physics
Separating signal from background
In collider experiments, rare signal events hide among far more common background. Classifiers trained on simulated events learn subtle combinations of measured quantities that separate them better than simple cuts.
Real exampleSearching for exotic particles with deep learning (Baldi, Sadowski and Whiteson, 2014)Showed on benchmark collider datasets that deep neural networks improved signal-versus-background classification compared with shallow machine learning models.
Fast emulation of expensive simulations
Detailed detector simulations are a major computing cost. Generative models learn to produce realistic simulated output much faster, for use where full precision is not needed.
Real exampleCaloGAN (Paganini, de Oliveira and Nachman, 2017)Used generative adversarial networks to simulate 3D particle showers in electromagnetic calorimeters as a faster alternative to full simulation.
Inferring parameters from experiments
When a simulator can produce data but its likelihood is intractable, neural networks can learn the relation between theory parameters and simulated observations, then apply it to real data.
Real exampleThe frontier of simulation-based inference (Cranmer, Brehmer and Louppe, 2020)Reviewed simulation-based inference and described how machine learning is giving new momentum to inference with complex scientific simulators.
Classifying detector noise
Sensitive instruments record transient noise that can mimic or hide real signals. Classifiers sort these glitches into types so their sources can be found and removed.
Real exampleGravity Spy (Zevin et al., 2017)Combined citizen science on the Zooniverse platform with machine learning to classify glitches in the LIGO gravitational-wave detectors.
Can physicists learn machine learning?
Yes, and physicists are among the best prepared. Linear algebra, calculus, statistics and programming already overlap heavily with machine learning; the main additions are machine learning practice, deep learning frameworks and validation habits.
- PythonMove analysis into Python; most physics data tools now have Python interfaces.Learn it in Python
- Statistics for machine learningLikelihoods, hypothesis tests and uncertainties connect classical analysis with model outputs.Learn it in Statistics
- Classical machine learningBoosted decision trees remain standard in many analyses and are strong baselines.Learn it in Machine Learning
- Deep learningGraph networks, transformers and generative models for detector data and simulation.Learn it in Deep Learning
- Physics-aware evaluationCheck calibration, systematic uncertainties and behaviour under differences between simulation and data.
Moving into machine learning for physics
If you already work in the field
Physicists bring mathematics, simulation experience and careful handling of uncertainty. Add machine learning practice through a project on open data, such as CERN Open Data or a public gravitational-wave or survey release, and learn PyTorch. Many physicists also move into machine learning research, data science or engineering roles in industry.
If you come from data science or ML
Machine learning practitioners should learn the physics of the target experiment and how its simulations and systematic uncertainties are produced. Collaborations publish tutorials and open datasets, and joining a physics group as a machine learning specialist is common.
Roles that do this workResearch physicistMachine learning scientistData scientistResearch software engineer
Tools you will use
| Tool | What you use it for |
|---|---|
| scikit-learn | Train classifiers and regression baselines on event and measurement tables. |
| PyTorch | Build neural networks for event classification and simulation emulators. |
| JAX | Write differentiable physics models and fit them with gradient-based optimisation. |
| Uproot | Read ROOT files from particle physics experiments into NumPy arrays. |
Public data to practise on
- HIGGS (UCI): Simulated collision events labelled as Higgs signal or background, with low-level and derived features.
- CERN Open Data Portal: Collision data, simulations and analysis examples released by LHC experiments.
Your first project
Download the UCI HIGGS dataset, classify signal and background events with gradient boosting, and compare it with logistic regression using ROC AUC, once on the raw features and once on the physicist-derived high-level features.
Common questions
Is machine learning used in physics research?
Yes, across particle physics, astrophysics, condensed matter and fluids, for classification, fast simulation, inference and control of experiments. Results are judged by the same statistical standards as other analyses.
Do physicists make good machine learning engineers?
Many do, because physics training covers linear algebra, optimisation, statistics and coding. The usual gaps are software engineering practice and production deployment, which can be learned on the job.
What is simulation-based inference?
It is a family of methods for inferring model parameters when a simulator can generate data but its likelihood cannot be written down. Neural networks learn the link between parameters and simulated data and are then applied to real observations.
How do physicists handle uncertainty in machine learning models?
By propagating systematic uncertainties through the model, calibrating outputs on control samples and testing on simulations with varied assumptions. A classifier score is treated like any other observable.
Courses for these skills
- Deep Learning with PyTorchUnderstand neural parameterisation and training.Syllabus available · lessons open as produced
- Maths for Machine LearningConnect linear algebra and variational objectives.26 published lessons · all planned lessons open
- Statistics and Probability for Machine LearningAssess sampling uncertainty.41 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.