This website is no longer updated.

As of 1.10.2022, the Faculty of Physics has been merged into the TUM School of Natural Sciences with the website For more information read Conversion of Websites.

de | en

Modern Deep Learning Learning for Physicists

Module NAT3008

This module handbook serves to describe contents, learning outcome, methods and examination type as well as linking to current dates for courses and module examination in the respective sections.

Basic Information

NAT3008 is a semester module in language at which is offered irregularly.

This Module is included in the following catalogues within the study programs in physics.

  • Specific catalogue of special courses for condensed matter physics
  • Specific catalogue of special courses for nuclear, particle, and astrophysics
  • Specific catalogue of special courses for Biophysics
  • Specific catalogue of special courses for Applied and Engineering Physics

If not stated otherwise for export to a non-physics program the student workload is given in the following table.

Total workloadContact hoursCredits (ECTS)
150 h 90 h 5 CP

Responsible coordinator of the module NAT3008 is Lukas Heinrich.

Content, Learning Outcome and Preconditions


Since its breakthrough moment in 2012, Deep Learning has made significant strides in all aspects of society including the natural sciences. This course will over a broad overview over Deep Learning techniques for inference, i.e. deep regression and classification, as well as for generative modelling. The course starts out with a short review of the pertinent elements of statistics, machine learning and array-based programming in which we aim to distill what aspects of deep learning differentiate it from classical machine learning. Following this refresher, we will discuss a range of problem specific architectures for various data modalities such as image data, graph data, sequence data and discuss components such as attention mechanisms, residual connections, gated units etc. We will then formulate classic ML tasks such as regression and classification and generative modelling in the language of probabilistic Machine Learning and discuss gradient-based optimization. Here you will learn how to train neural networks practically. Throughout the course we will discuss how these methods are applied in fundamental physics research.

Learning Outcome

After successful completion of the module the students are able to:

  1. Recall necessary Statistics, Learning Theory, and Array Programming
  2. Understand the Benefits of Deep Learning and Overparametrization with respect to other approaches to ML and AI.
  3. Have a Broad Overview over deep neural networks architectures (Recurrent Networks, Convolutional Networks, Attention, Transformers, Graph Neural Networks)
  4. Understand Generative Models (GANs, VAEs, Normalizing Flows, Diffusion Models, …)
  5. Understand Deep Learning Inference as Amortized Bayesian Analysis
  6. Use Differentiable Programming and Gradient-based Optimization
  7. Have an Overview over Machine Learning Frameworks
  8. Have an Overview over Applications in fundamental physics Research
  9. How to train deep neural networks for a variety of data modalities (image, sequences, structured data)


No preconditions in addition to the requirements for the Master’s program in Physics.

Courses, Learning and Teaching Methods and Literature

Courses and Schedule

VO 2 Modern Deep Learning Learning for Physicists Heinrich, L. Fri, 13:00–15:00, ZEI 0001
UE 2 Exercise to Modern Deep Learning Learning for Physicists
Responsible/Coordination: Heinrich, L.
dates in groups

Learning and Teaching Methods

The course will taught using prepared lectures with intermittent periods of discussion to reflect on the material. Where applicable, we will look at live code examples and interactive visualizations in order to deepen the understanding of the discussed concepts. The various deep learning methods will then be connected to real-world applications from fundamental physics research


The material used in the lecture includes * slides * Jupyter Notebooks * Interactive Visualizations and will be made available publicly


no info

Module Exam

Description of exams and course work

There will be a written exam of 90 minutes duration. Therein the achievement of the competencies given in section learning outcome is tested exemplarily at least to the given cognition level using comprehension questions and sample calculations.

For example an assignment in the exam might be:

  • Discuss how Deep Learning differentiates itself from other Types of Approaches to AI
  • What's the Bias-Variance Tradeoff?

Participation in the exercise classes is strongly recommended since the exercises prepare for the problems of the exam and rehearse the specific competencies.

Exam Repetition

The exam may be repeated at the end of the semester.

Top of page