Deep Learning and Materials Discovery, introduction and methods
Lecturer: Maxime Moreaud
In a few lines, what will be covered during the course:
- demystify deep learning
- training convolutional neural networks, with several main applications: classification, segmentation, noise reduction
- using trained network with plug im!
This workshop deals with artificial intelligence and its application to image processing, focusing on classification, segmentation and noise reduction. The aim is to demonstrate the effectiveness of deep convolutional neural networks in these areas, while providing participants with the skills needed to apply them autonomously.
We begin with a presentation of deep learning highlighting the use of deep convolutional neural networks of the encoder-decoder type.
Participants will set up a supervised training process using a stratified patch procedure, witch allows efficient learning even with a small number of images. The entire course will be taught in Python, using Jupyter Notebook.
Participants will then learn how to use the networks trained with the plug im! platform ( https://www.plugim.fr ) to process new images quickly and efficiently.
Detailed instructions on how to install the necessary Python environments, as well as source code and example data, will be made available a few days before the course.
This course provides comprehensive hands-on experience of deep learning applied to classification, semantic segmentation and image denoising. Thanks to a hands-on approach, participants will acquire applicable skills and be able to directly test the performance of deep convolutional neural networks, while benefiting from the efficiency of the plug im! platform for their future image processing projects.
Maxime Moreaud holds a PhD in mathematical morphology from the Ecole des Mines de Paris (France) and, since 2017, an accreditation to supervise research (French HdR) from the University of Lyon (France). Since 2008, he has been a researcher at IFP Energies nouvelles (IFPEN), a French research centre specialising in the fields of energy, transport and the environment. He has been a research project leader in imaging, then in digital materials and deep learning, leading multidisciplinary teams and supervising more than 15 doctoral theses. He has also worked closely with the Centre for Mathematical Morphology at the Ecole des Mines de Paris, as a research associate from 2016 to 2022. He has also been welcomed as a visiting researcher in 2021 and 2022 at the CERVO research centre at Laval University in Quebec, Canada, to initiate an ambitious roadmap on the contribution of artifical intelligence (AI) to neurophotonics. From 2024, he will oversee the implementation of fundamental research in AI applied to the various research and innovation activities ot IFPEN.
Learning Atomic Interactions
Lecturer: Ralf Drautz
Atomic interactions are important for many problems in materials science, chemistry, physics and biology. In my lecture I will first introduce the concept of interacting atoms and briefly discuss its origin in quantum mechanics. I will then highlight the relevance of atomic interactions for predictive simulations in materials science and chemistry and motivate why accurate and efficient models of atomic interactions are required. This will be followed by a short summary of the history of interatomic potentials and the main characteristics of empirical or classical potentials. More than 20 years ago the first machine learning -or machine learned - potentials (MLPs) were developed and I will give an overview of the most important MLPs. Next I will take a step back and analyse the underlying structure of classical potentials and MLPs. This allows one to combine most of the available potentials in a general framework with clear inputs and structure. It further enables systematic extensions for other degrees of freedom, for example, magnetic moments and charges. Finally, the lectures will be completed with a demonstration and hands-on tutorial for fitting a MLP.
Processing experimental data
Lecturer: Jaysen Nelayah
Over recent decades, electron microscopy has rapidly evolved as a major investigating tool in material science. Through major technological breakthroughs and novel methodologies in data acquisition and analyses, EM nowadays enables access to the atomic and chemical structure of nanomaterials down to the atomic scale, paving the way to an improved understanding of nanomaterials and ultimately, to their rational design for applications across various field such as environment, energy and biomedicine. Alongside, this success has brought its share of bottlenecks like high throughput 2D and 3D data acquisition at rates which exceeds those at which they can be analysed, or possibility to study nanomaterials liquid in gas environments but at the expense of degraded image contrasts. Over the very recent years, AI has been identified as a ground-breaking tool to answer such conundrum.
In this course, we will touch on both the main EM techniques and the practical aspects of EM data acquisition and analysis, moving from traditional practices to the implementation of deep-learning strategies. We will cover the peculiar attributes of EM data structure and present recent deep-learning based developments for EM image/spectrum denoising, segmentation, classification. We will also reflect on beyond state-of-the-art methodologies for the characterization of nanomaterial structure in an electron microscope opened by deep-learning. Finally, we will discuss the adaptability of these methodologies to various microscopies and samples.
Dr. Jaysen Nelayah is an associate professor in experimental physics from the Materials and Quantum Phenomena laboratory at Université Paris Cité. His research covers the physics of metallic nanoparticles and aims at resolving the intricacies between the structural and physico-chemical properties of these systems using high energy electrons in a transmission electron microscope. His current research interest explores in operando the dynamics of the structural and thermodynamics of metal-gas interfaces in metallic nanoparticles. He has recently extended his research interest to advancing electron microscopy using deep learning.
