Intelligent Physics · Module Presentation

A new master's module bringing data-driven methods into the materials physics curriculum — a course of this kind does not yet exist elsewhere in the Algerian university system.

Module Launch Master's Program Materials Physics First in Algeria

A new module, and a new kind of question

Materials physics has traditionally been taught and practiced as a theory-first discipline: a Hamiltonian is written down, a structure is relaxed, an electronic structure is solved, and a property is extracted from first principles. This approach remains the foundation of the field and will continue to be. But over the past decade, an entirely different way of asking questions about materials has matured alongside it — one that starts not from the Schrรถdinger equation but from data, and asks what statistical structure can be learned from the accumulated results of decades of simulation and experiment.

Intelligent Physics: Machine Learning Methods for Materials Science is a new master's-level module built specifically to teach this second way of asking questions, without abandoning the first. It is, to the best of available knowledge, the first dedicated module of its kind offered within an Algerian university physics curriculum.

14Weeks
6ECTS
1stOf its kind in Algeria
M1/M2Materials Physics

Why this module, and why now

Materials discovery increasingly depends on the ability to search across vast compositional and structural spaces that cannot be explored by first-principles calculation alone, given realistic computational budgets. Internationally, the response to this constraint has been the rapid growth of materials informatics: structure-property datasets numbering in the hundreds of thousands, graph-based architectures purpose-built for crystal structures, and active-learning pipelines that pair machine learning surrogates directly with ab initio codes. Algerian physics master's programs have, until now, had no dedicated module preparing students to participate in this shift.

This module addresses that gap directly, and deliberately does so without treating machine learning as a replacement for ab initio simulation. The course is built around codes and workflows already familiar to ab initio practitioners — Quantum ESPRESSO, VASP, WIEN2k — and treats machine learning as an additional instrument in the same toolbox, governed by the same standards of validation and physical reasoning that any computational method in physics must meet.

What this module is not

This is not a general-purpose data science or coding bootcamp. Every method introduced is anchored to a physical question in materials science, and every dataset used is drawn from, or structured like, real materials property data. Students leave with the ability to build and critically evaluate a machine learning pipeline for a materials problem — not a generic familiarity with popular algorithms.

What the module covers

Across fourteen weeks, the module progresses from statistical learning foundations through classical and ensemble methods, into the specific problem of representing crystal structures and molecules numerically, then into neural network and graph neural network architectures built for structure-property prediction, and finally into uncertainty quantification and the integration of machine learning surrogates directly into ab initio workflows.

  • Statistical foundations: bias-variance trade-off, cross-validation, honest generalization assessment
  • Regression, classification, and ensemble methods (Ridge, LASSO, Random Forest, XGBoost) applied to materials property prediction
  • Unsupervised learning for compound family discovery and phase mapping
  • Structural representation: descriptors via ASE, DScribe, and MatMiner
  • Neural networks, convolutional architectures for spectroscopic data, and graph neural networks (CGCNN, ALIGNN) for direct structure-to-property learning
  • Uncertainty quantification, physics-informed constraints, and active-learning loops coupling machine learning surrogates to ab initio calculations
  • Interpretability, reproducibility, and the major open materials databases (Materials Project, AFLOW, OQMD)

The module closes with an independent student project, in which each student builds and defends a small original machine learning pipeline addressing a real materials-property question — with the explicit option to connect the project to ongoing thesis research.

Who this module is for

The module is designed for master's students in materials physics, with prerequisites limited to a standard condensed matter physics background, introductory quantum mechanics, basic statistics, and elementary programming experience. No prior machine learning background is assumed. Students already engaged in ab initio research — whether on Heusler alloys, MAX phases, transition-metal chalcogenides, or related compound families — will find direct points of contact between the module's methods and their own thesis work throughout the semester.

About this series

This blog will publish one post per week of the module, following the course sequence from Week 1 through Week 14, along with supporting material, lab guidance, and discussion of open questions in machine learning for materials science as they arise during the semester. The next post introduces the foundational question of the course: why a data-driven approach to materials physics complements, rather than competes with, the first-principles tradition the field has built over the past century.

Intelligent Physics — Master's Module Week 1 →

INTELLIGENT PHYSICS: MACHINE LEARNING METHODS FOR MATERIALS SCIENCE — MASTER'S MODULE