From Data to Models: An Introduction to Machine Learning and Artificial Neural Networks for Scientific Discovery

How can we turn data into predictive models, and how can physics principles guide what these models learn? This talk provides an accessible introduction to data science, machine learning, and neural networks for a multidisciplinary scientific audience. Beginning with the role of data in scientific discovery, we discuss how to formulate useful questions, select appropriate data, and evaluate predictions on unseen examples. We then introduce the principles of machine learning and representative neural network architectures, including convolutional networks, autoencoders, and recurrent networks, highlighting their roles in pattern recognition, dimensionality reduction, and temporal modeling. Particular attention is given to generalization, overfitting, and interpretability. Finally, we explore how governing equations and mechanical structure can be incorporated into learning through physics-informed, Hamiltonian, and Lagrangian neural networks, using pendulum dynamics as an illustrative example. The emphasis is on conceptual understanding and informed modeling choices, without requiring prior expertise in machine learning.