Familiarity with Machine Learning Foundations (or equivalent knowledge of supervised and unsupervised learning, neural networks, and optimisation methods) is assumed. Students are also expected to possess background knowledge in astrophysics, cosmology, and particle physics broadly equivalent to that acquired by an Oxford Physics student by the end of the third undergraduate year.
Lecturers: Deaglan Bartlett, Andrei Constantin, Adrian Oeftiger, Holly Pacey
This course is intended as a natural continuation of Machine Learning Foundations for Mathematics and Physics. Whereas the foundations course introduces the mathematical and algorithmic principles underlying modern machine learning methods, together with applications of these methods to problem solving in mathematics and theoretical physics, the present course focuses more specifically on how machine learning can be used in contemporary scientific research to bridge the gap between experimental or observational data and theoretical understanding. In particular, the course emphasises the use of machine learning methods to extract physical information from complex data, perform statistical inference, construct effective models, and assist in the discovery of new theoretical structure.
A central aim of theoretical physics is the construction of predictive models and governing principles from empirical observation. This course explores how modern machine learning methods can be used not merely for prediction, but as tools for scientific inference, probabilistic modelling, model selection, optimisation, and theory discovery. The course focuses on methodologies which allow physical information to be extracted from high-dimensional and computationally complex datasets, including simulation-based inference, surrogate modelling, uncertainty quantification, representation learning, generative modelling, anomaly detection, causal discovery, symbolic regression, and reinforcement learning. Throughout, emphasis is placed on understanding the common methodological principles underlying machine learning approaches across different areas of physics, and on identifying the physical assumptions which make these approaches effective.
Applications and case studies will be drawn from astrophysics, cosmology, particle physics, accelerator physics, condensed matter physics, and theoretical physics. These may include emulators for cosmological summary statistics and accelerated N-body simulations; likelihood-free parameter inference for cosmological and particle-physics models; particle identification and collision-event classification at colliders; anomaly detection for rare-event searches; machine-learning methods for jet physics and detector reconstruction; dimensionality reduction and latent-space representations in many-body systems; surrogate models for accelerator optimisation and uncertainty quantification; generative models for physical simulations; symbolic regression for discovering effective dynamical laws.
- Scientific inference and the philosophy of data-driven discovery
- Probabilistic modelling and Bayesian parameter inference
- Simulation-based inference and likelihood-free methods
- Neural posterior and ratio estimation
- Surrogate modelling and emulation
- Gaussian processes and uncertainty quantification
- Bayesian optimisation and experimental design
- Representation learning and physically informed latent spaces
- Data compression and summary statistics
- Generative models for scientific simulation
- Classification, anomaly and novelty detection
- Causal inference and discovery from observational data
- Symbolic regression and interpretable model discovery
- Reinforcement learning for control and optimisation
- Foundation models and large-scale scientific models
- Open problems in AI-assisted scientific discovery