Machine Learning for Materials Discovery
Machine Learning in an open science setting is a place to optimize future material development through screening and prediction of material-properties. Predictions are based on known properties of materials at all length scales. In developing new materials, machine learning will optimize the selection process and allow for a more effective development process
Read more about open science machine learning here: https://spoman-os.org/sustainable-polymersmachine-learning/
When “what” is more important than “why” - applications of Machine Learning for classification by Brian Vinter, Computer Science, Copenhagen University
The traditional scientific model development process takes observations and from those models are build that may predict future observations and may be used to detect observational outliers. Using deep-learning, one may build a black-box model directly from observations. Such a black-box model may also predict future observations and efficiently identify outliers. Black-box models provides no deeper understanding of why a system behaves as it does, but may still be used as a supplement to conventional data-analysis. In this talk we will introduce the difference between conventional data-analysis and deep learning, and will introduce applications from optic-light, x-ray and climate research.
Programming common sense: Utilising domain knowledge and common-sense spatial reasoning with applications in architectural design by Carl Schultz, Department of Electrical and Computer Engineering (ECE), Aarhus University
While there has been tremendous progress in deep learning methods, particularly in computer vision, such methods are not well suited to common-sense analysis and tasks that rely heavily on everyday semantic knowledge about the way the world works, e.g.: “The trophy would not fit in the brown suitcase because it was too big. What was too big? (a) the trophy or (b) the suitcase?” (from the Winograd schema challenge) In this presentation I will discuss these interesting limitations, and our research into combining machine learning with common-sense spatial reasoning and “knowledge representation and reasoning” (within Artificial Intelligence), that aims to formally capture the way that humans conceptualise and reason about objects and events. Applications include qualitative analysis, explanation, and Question/Answering for architectural design and video scene interpretation.
Data Leak Prevention - Identifying sensitive information in text using deep learning by Ira Assent, Dept. Computer Science, Aarhus University
Data Leak Prevention, detecting sensitive information in documents prior to publication, is increasingly important in industry and government. A core challenge are vague and complex definitions of sensitivity, as well as the mix of sensitive and non-sensitive content in the same document. We present an approach that extracts sub-structures from sentences using methods from Natural Language Processing. We build a Recursive Neural Network model to learn sensitivity scores for these sub-structures, and to identify sensitive documents.
Identifying optimal structure of molecules and solids byBjørk Hammer, Dept. of Physics and Astronomy, Aarhus University
The search for new materials may be done using quantum mechanical computer codes that establish structure-property relations. Traditionally, the codes have been operated manually, but improved algorithms and computational capabilities allow for automated search setups. In the talk, it is demonstrated how using machine learning analysis of all produced structure-property relations allows for faster identification of the most stable structures of molecules, nano-particles, and solid surfaces.
COMSOL Multiphysics Simulation Package by Kristian Ejlebjærg Jensen, COMSOL
COMSOL Multiphysics is a fully featured simulation package with a wide range of solvers as well as functionality for pre- and post-processing. We employ a white-box approach to simulation, which allows users to see and change the weak form of the equations. Our optimization module supports parameter-, shape- and topology optimization for arbitrary control variables and objectives. Topology optimization has the most design freedom, and it allows for the identification of optimal 3D shapes without a priori assumptions about the geometry. I will show examples from structural compliance minimization and reactive flows. Issues within the field relate to physics specific interpolations, numerical parameters and local minima. Any ideas to resolve these issues are welcome and COMSOL‘s livelink for MATLAB can be used as a platform for testing out ML and AI methods in the context of topology optimization or simulation in general. The resulting knowledge can be condensed into a simulation app with a simple GUI, thus democratizing state-of-the-art computational engineering.
Topology Optimization for Electromagnetic Field Enhancement and Upconversion Applications by Søren Peder Madsen, Dept. of Engineering, Aarhus University
SunTune – High-efficiency Solar cells by spectral transformation using nano-optical enhancement – is a 4-year research project headed by Aarhus University and sponsored by Innovation Fund Denmark. DTU and SDU are among the partners in the project. The project will boost the efficiency of solar cells by converting two low energy (long wavelength) photons into one photon with energy above the Si band gap, which then adds to the current production in the Si cell. This is not efficient at natural sunlight intensities, therefore the light needs to be focused to enhance the non-linear upconversion process. Metal nanoparticles on top of an erbium-doped TiO2 upconverting thin film can be used to focus the part of the incoming light that matches the absorption band of erbium into the thin film. The focusing efficiency depend on the geometry of the metal nanoparticles and gradient-based topology optimization is used to efficiently calculate optimized designs. Topology optimization allows for almost unlimited freedom in the design and several techniques are used to make the design production-friendly with respect to electron beam lithography. Physically, the optimization process can exploit scattering, plasmonic, and waveguiding as well as diffraction effects for periodic arrays of metal nanoparticles. Designs utilizing the latter tend to show very good performance but are also extremely sensitive with respect to changes, e.g. in wavelength and angle of incidence. Topology optimization can find non-trivial designs which are efficient and, at the same time, only no so sensitive to variations in wavelength, angle of incidence, and particle geometry.