ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING IN PHYSICS

General Foundation

The course is also offered to exchange programme students.

Upon successful completion of the course, the student should be able to: - Explain the basic principles and evaluate the advantages and limitations of the artificial intelligence approach to solving problems related to the subject of Physics and beyond. - Be able to address problems of applying machine learning both supervised and unsupervised. - Be able to apply these machine learning methods to different scenarios and problems related to the subject of Physics and beyond. - Be able to choose the best machine learning method for solving a specific physical problem based on the characteristics of the problem and the available data. - Compare and contrast alternative machine learning algorithms that are often used to solve optimization problems. - Distinguish the types of neural networks, in terms of the data they can process and the problems they are used to solve.

Course Content Introduction to Artificial Intelligence and Machine Learning: Definitions and basic concepts (AI, ML, Deep Learning). Learning categories. Machine learning applications - limitations. Introduction to Python and basic libraries. Supervised Learning: Performance criteria (accuracy, precision, recall, ROC) - model comparison. Classification methods (k nearest neighbors, decision trees). Practical application in predicting natural variables from experimental or simulated data. Artificial Neural Networks and Training: Basic principles (Perceptron, logistic regression, softmax function). Common activation functions (ReLU, tanh, sigmoid). Multi-layer neural networks. Deep Learning Models. Recursive and Convolutional Neural Networks – Applications in Physics: Recursive neural networks (CNN, RNN, LSTM). Time series analysis of physical systems. Phase recognition of matter using CNN. Chaos and prediction using LSTM. Overview of AI Applications in Physics: Applications in astronomy, high-energy physics, chaotic systems, and solid-state physics. AI for accelerating simulations and solving inversely defined problems. Analyzing experimental data with machine learning tools. Physics-Informed Neural Networks (PINNs): Incorporating physical laws into the training process of neural networks. Differential equations and PINNs. Examples of solving PDEs (e.g. heat equation, wave equation). Using automatic differentiation. Analysis of Complex and Nonlinear Systems. Reconstruction of Attractors (Takens), use of entropy and finding the Lyapunov exponents of nonlinear systems. SINDy and Symbolic Regression for extracting equations from data. Applications to nonlinear dynamical systems. Reinforcement Learning: Basic principles of Reinforcement Learning (RL). Environments - State, Action, Reward. Policy Optimization, Learning Q-values. Applications of RL to particle systems and robotic vehicle navigation. Reservoir Computing. Definition - Basic principles - Advantages. Basic reservoir computing models: Echo State Networks (ESN), Liquid State Machines, Time Delay Reservoirs. Reservoir computer architectures and their physical implementations (optical, circuit, mechanical Reservoir). Application of Reservoir Computing in time series forecasting and pattern recognition/categorization.

Artificial Intelligence, Machine Learning, Neural Networks, Physics-Informed Neural Networks, Reservoir Computing

22-07-2025