Description
This training course enables participants to understand and apply Artificial Intelligence (AI) tools to regression and optimal control tasks in physics. Fundamental AI concepts are first introduced theoretically, then systematically implemented in Python notebooks for hands-on practice using libraries such as ScikitLearn, Keras/TensorFlow, and DeepXDE. Major methods from Deep Learning and Deep Reinforcement Learning will be covered (MLP, CNN, autoencoders, VAE, PINN, etc.) and tested on physics-related datasets (rocket engine, transient heat equation, point mechanics). Specific aspects of AI applied to physics will also be discussed (meshes, physics-informed networks, etc.).
🎯 Training objective
Use open-source libraries (ScikitLearn, Keras/TensorFlow, DeepXDE) to train neural networks on regression and optimal control problems in physics.
✅ Learning outcomes
By the end of this training, participants will be able to:
-
Understand the fundamental concepts of AI
-
Formulate a learning problem (regression/control, cost function, hyperparameter choices, etc.)
-
Use ScikitLearn and Keras/TensorFlow for regression-type problems, and DeepXDE for PINNs and inverse problems
-
Evaluate and test neural network training results
📚 Teaching methods
The training alternates between lectures and practical sessions. A final multiple-choice test is used for evaluation. Training rooms are equipped with computers, and participants may work in pairs.
👤 Lead instructor: Michaël Bauerheim
👥 Target audience
This course is designed for physicists (engineers, PhD students, post-docs, interns) wishing to work with modern AI tools on physics data.
🔑 Prerequisites
Participants must:
-
Be employed by a European company (employer certificate required)
-
Hold a Master’s degree (Bac+5 or equivalent) or higher
-
Know the basics of Python
-
Have general knowledge in physics and mathematics
-
Have at least a B2 level in English (CEFR), as the course may be taught in either English or French depending on the audience
To confirm prerequisites, applicants must complete preliminary questionnaires and achieve at least 75% correct answers to be eligible.
📝 Registration
I certify that I obtained at least 75% correct answers — I register.
📅 Registration deadline: 15 days before the start of the course
Before registering, please inform us of any specific constraints (schedule, health, availability, etc.) by email: training@cerfacs.fr
📆 Program
Schedule: 9:00 AM – 5:00 PM (1-hour lunch break)
Day 1
Introduction, regression problems, overfitting/underfitting, train/validation/test sets, regularization, introduction to ScikitLearn, rocket engine application.
Day 2
Introduction to deep learning: MLP, backpropagation, optimization (SGD, RMSprop, ADAM), MLP playground test, introduction to Keras/TensorFlow.
Day 3
Introduction to CNNs, padding and resolution, autoencoders and VAEs, overview of generative AI, CNN implementation for the transient heat equation.
Day 4
Introduction to Physics-Informed Neural Networks (PINNs), theory and backpropagation, introduction to DeepXDE. Applications of PINNs to the transient heat equation and inverse problems with DeepXDE.
Day 5
Introduction to reinforcement learning (RL): Bellman equation, Q-learning, RL applications in physics. Hands-on: RL applied to a simple point mechanics problem. Final exam (MCQ) and Q&A session.
📊 Evaluation
A final multiple-choice exam will take place at the end of the training.