STORMING will develop breakthrough and innovative structured reactors heated using renewable electricity, to convert fossil and renewable CH4 into CO2-free H2 and highly valuable carbon nanomaterials for battery applications. More specifically, innovative Fe based catalysts, highly active and easily regenerable by waste-free processes, will be developed through a smart rational catalyst design protocol, which combines theoretical (Density Functional Theory and Molecular Dynamics Calculations) and experimental (cluster) studies, all of them assisted by in situ & operando characterisation and Machine Learning tools. The electrification (microwave or joule-heated) of structured reactors, designed by Computational Fluid Dynamics and prepared by 3D printing, will enable an accurate thermal control resulting in high energy efficiency. The project will validate, at TRL 5, the most promising catalytic technology (chosen considering technological, economic, and environmental assessments) to produce H2 with energy efficiency (> 60%), net-zero emissions, and decreasing (ca. 10 %) the costs in comparison with the conventional process. The dissemination and communication of the results will boost the social acceptance of the H2-related technologies and the stakeholder engagement targeting short-term process exploitation and deployment. The key to reach the challenging objectives of STORMING is the highly complementary and interdisciplinary consortium, where basic and applied science merge with engineering, computer and social sciences.

Our company has developed advanced chemical imaging capabilities which we offer as a service to industry, helping our clients accelerate their R&D. Our imaging approaches yield rich and large datasets that contain an abundance of physico-chemical information. This project will use artificial intelligence approaches to reconstruct X-ray scatter-based chemical tomography data from large objects.Large objects pose a problem due to geometric blurring of the scattered signals on the receiving detector, preventing conventional reconstruction approaches. We have spent considerable resources developing a non-linear least-squares algorithm to address this but it is computationally demanding and because of this imposes resolution limits on the reconstructed data (i.e. small images size). We have realised though that the problem has several features which indicate that it can be tackled by using deep learning approaches. Additionally, we have the ability to generate very large simulated labelled datasets that can be used as training sets for supervised learning using convolutional neural networks (CNNs). This is in addition the very large real data sets we have at our disposal. Whilst there are existing attempts to reconstruct conventional tomography data using CNNs, we are planning to develop new CNNs for reconstructing chemical (hyperspectral) tomography data and indeed overcome the parallax problem. The project thus is innovative both in terms of approach and application and will push the opportunities in this emerging field.

Through this project we seek to identify and develop methods to automatically segment complex chemical imaging datasets, e.g. XRD-CT, Raman mapping and other hyperspectral imaging techniques. The goal is to identify the minimum number of unique chemical environments in a dataset, without needing prior knowledge or input as to the expected identity or number of components present. The output from this segmentation will then be used to inform subsequent data quantification and analysis steps, and also to see if it is possible to identify correlations between each of these components. A successful outcome will be of benefit to a broad range of industry and chemical services companies, as many analytical methods suffer from similar segmentation challenges.

This project will use machine learning approaches to extract physico-chemical information from chemical imaging data. This novel approach will tackle an emerging problem in this field, namely how to automatically identify and extract chemical signals from the rich and ever-larger datasets that it is now possible to collect. There are several features that suggest this problem can be tackled using machine learning approaches. We have developed software for the rapid simulation of chemical imaging data, and we can use this to generate large labelled datasets for training the convolutional neural networks (CNN) that we will build. In addition we have substantial libraries of real data which the developed CNN's can be tested against.

no public description