Machine Learning for
Astroparticle Physics
A Crash-Course in Simulation-Based Inference
ICTP-SAIFR SAMMA 2026 PhD School — São Paulo, 2026 — Christoph Weniger (GRAPPA / University of Amsterdam)
-
Lecture 1aMachine Learning in Astroparticle PhysicsA guided tour of where ML has actually landed in APP/GW research, organised around the scientific-method loop: theory, simulator, instrument, data, inference. Closes on the two ways to do Bayesian inference with a simulator — setting up Lec 1b.Overview DINGO IceCube LIGO Emulators SBI taxonomy
-
Lecture 1bApproximate Bayesian ComputationA three-demo staircase: rejection ABC → add a summary statistic → KDE in (θ, x)-space. Live, slider-driven demos expose the two failures — summary identification and high-D density estimation — that motivate the rest of the school.ABC Rejection sampling Summary statistics KDE Curse of dimensionality
-
Lecture 2aLinear Regression as Conditional Density EstimationLeast-squares fitting is conditional density estimation: the curve is μ(x), the residual scatter is σ, together giving q(θ | x) = 𝒩(μ(x), σ²). Maximum likelihood, the closed-form solution, and where a fixed basis starts to overfit.Conditional densities Basis functions Maximum likelihood Closed-form solution Overfitting Validation
-
Lecture 2bDeep Neural Networks for Density EstimationThe hand-picked basis was the bottleneck — so learn it. Neurons as learnable basis functions, the universal approximation theorem, training by gradient descent and mini-batch SGD, and how stacking layers builds richer effective bases.Neural networks Activations Universal approximation Gradient descent SGD Depth
Session 1A — Train & Dissect an MLP (exercises)
Session 1A — solutions
Session 1B — Gravitational Wave NPE (exercises)
Session 1B — solutions
-
Companion to Lecture 2b OptionalMaking Deep Learning Work in PracticeA self-contained, orthogonal deep-dive into actually training deep networks: backpropagation, vanishing/exploding gradients and initialisation, normalisation, momentum and Adam, learning-rate schedules, and why overparameterised models still generalise (double descent, spectral bias). Not part of the core SBI thread — dip in for the training mechanics behind Lecture 2b.Backpropagation Gradient flow Initialisation Normalisation Adam Double descent Spectral bias