PhD Student – Physical modelling via deep neural networks
We are currently accepting applications for the above mentioned position. This is a unique opportunity for highly motivated students, recently graduated from the university in Physics or related fields, to join one of DIPC’s high-profile research teams.
Physical modelling via generative low-complexity deep neural networks with position-dependent input.
A PhD student is being sought to carry out a project in Deep Neural Networks (DNNs), based in the Dept. of Polymers and Advanced Materials: Physics, Chemistry and Technology, Faculty of Chemistry, University of the Basque Country, San Sebastian, Spain (UPV/EHU) in active collaboration with Dept. of Mechanical Engineering, Jiangsu Key Laboratory for Design and Manufacture of Micro-Nano Biomedical Instruments, Southeast University, Nanjing, China (SEU).
Most importantly, after training, the simple DNN can be used out-of-the-box to generate new segmented images in a pixel-by-pixel manner for unknown experimental conditions (generalisation, i.e. interpolation and extrapolation). This generative functionality is remarkable considering the simplicity of the approach (an uncomplicated DNN with spatial-dependent input). Previously, such functionality has been restricted to highly specialised Convolutional Neural Networks (CNNs) with millions of trainable parameters, such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs). The low complexity of the DNN enables simple/fast modifications/adaptations (e.g. for other projects or in other fields) as well as the use of less-sophisticated computational resources during both training and execution, leading to moderation in power consumption (an increasingly important aspect) while simultaneously simplifying model deployment on mobile devices with limited resources (wherever this aspect may become relevant).
DESIRED BACKGROUND & COMPETENCES
The successful candidate is expected to perform a number of tasks, including (A) physical modelling of various systems, (B) automated image segmentation, (C) optimization of the network structure, (D) improved extrapolation. Previous experience in DNNs is required. In particular, direct experience in the use of one or several open-source software libraries is a must (TensorFlow, PyTorch, Keras, Caffe, etc…). TensorFlow is favoured, since the already existing code is based on it. Experience in other Machine learning (ML) techniques, such as support vector machines (SVMs), decision trees (DTs), etc… will be appreciated but secondary.
Experience/interest in Physics/Mechanical engineering/Numerical modelling will be highly appreciated. We are trying to build physically meaningful models, one for each of various physical systems, so that we can predict their behaviour under unseen experimental conditions. Previous experience in image segmentation, either by traditional methods (level set, etc…) or by current DNNs, will be highly appreciated. Experience/interest in extrapolation techniques will be highly appreciated as well. The ability of neural networks in general to generalise outside the range where they were trained is currently very weak. The use of image segmentation provides us with a privileged position, since it allows performing a numerical analysis of the dependence of the size and/or shape of the regions (number of pixels, area, centre of mass, typical width, moment of inertia, etc...) with respect to the experimental variables. This can be used to achieve an improved level of extrapolation outside the training range, which CNNs cannot provide (at least yet). Essentially, the difference here is that between data-driven learning (provided by CNNs) and data-driven discovery of underlying physical behaviour (our goal).
Finally, the candidate should demonstrate excellent/advanced proficiency in both written and spoken English to ensure a smooth collaboration with the Chinese group. For exceptionally bright candidates with a strong motivation and fast learning/working abilities, some of the previous requirements might be relaxed provided they can show convincing proof of such excellent skills (e.g. support letters). We believe in equality. If you are interested and you feel part of a social/scientific minority we encourage you to contact us.
- Contract duration: 1 year (possibility to extend up to 3 years)
- Target start date: 01/07/2023
We provide a highly stimulating research environment, and unique professional career development opportunities.
We offer and promote a diverse and inclusive environment and welcomes applicants regardless of age, disability, gender, nationality, ethnicity, religion, sexual orientation or gender identity.
DIPC is a research center whose mission is to perform and catalyze cutting-edge research in physics and related disciplines, as well as to convey scientific culture to society. Located in Donostia / San Sebastian (Basque Country, Spain), DIPC was born from a strategic alliance of both public institutions and private companies. Since 2008 DIPC is a 'Basque Excellence Research Center' (BERC) recognized by the Basque Government's Department of Education. In 2019, DIPC was also recognized as a "Severo Ochoa" Excellence Center by the Spanish Research Agency.
ABOUT THE TEAM
The PhD project is a natural continuation of a recent collaboration between Dr. M. A. Gosalvez (UPV/EHU) and Prof. Y. Xing (SEU). Gosalvez and Xing have worked jointly for approximately 17 years, as reflected by at least 20 co-authored publications in peer reviewed journals and 15 contributions in international conferences. Xing’s group is devoted to both experimental and computational work while Gosalvez’s group is dedicated to theoretical and computational modelling. While based in Gosalvez’s group, the PhD student will have the chance to visit Xing’s group twice during 2023-2024.
The project is based on a recent finding by the two groups. For physical systems exhibiting a strong spatial dependence (usually as a function of a few relevant experimental parameters) and when the number of images that display the spatial dependence is very reduced (due to excessive cost/time-consumption/legal-constraints/etc), one can directly use the spatial dependence to segment the images into physically meaningful regions, whereafter every pixel from every image will belong to one region (or class/label), thus resulting in a truly large pixel-based dataset with millions of classified pixels, even when obtained from a few images. The key idea is to use this simple piece of information (the region/class/label) as the output for two equally simple inputs: the experimental conditions and the pixel’s location within the image. In this manner, a low-complexity DNN (with a small number of hidden layers and neurons) can be easily trained into understanding the spatial dependence of the system as a function of the experimental parameters.