In its third funding round for Helmholtz AI projects, Helmholtz is investing 4 million euros in collaborative research projects in the field of applied artificial intelligence and machine learning.
Helmholtz AI strengthens the application and development of applied artificial intelligence (AI) and machine learning (ML). In the third call for projects of Helmholtz AI, a top-class international panel of experts selected 10 collaborative high-risk, high-reward research projects. Helmholtz is investing a total of 4 million euros in these novel approaches – half of which is provided by the Association’s Initiative and Networking Fund.
“I am delighted that numerous exciting and interesting project proposals were submitted again this year. Thanks to our international expert panel, we have identified the ten most innovative and promising projects with diverse applications of AI methods that address major challenges in various research areas and across Helmholtz’ research fields”, says Otmar D. Wiestler, the President of the Helmholtz Association. “A very active research community is visible here, which has found a catalyst in the Helmholtz AI platform that makes it possible to motivate and bring together innovative research design, state-of-the-art methods and excellent researchers.”
33 project proposals were submitted in the third Helmholtz AI project call. The 10 selected projects use novel analytical tools to solve pressing scientific and distinct transfer challenges using AI. The projects promote the testing of these new approaches, they are supported by several partners and will be implemented in up to three years – with the potential to quickly spawn larger follow-up projects.
“The Helmholtz AI projects are central to what we wish to achieve: spreading the opportunities of applying AI to unique datasets, building bridges between Helmholtz Centers and most importantly connecting researchers,” says Fabian Theis, the scientific director of Helmholtz AI. “This third cohort of projects expands the network between domain and method specialists and I am excited to see our open and dynamic community growing further.”
With the Helmholtz AI platform, Helmholtz is investing up to 11 million euros annually in six innovative research units and a transdisciplinary network for applied AI. This network builds on the community's strengths in AI and its unique data sets. In this way, Helmholtz AI contributes to solving the grand challenges of our time, fostering cross-field and cross-center collaboration, also with external university and industry partners. The network keeps growing – the next call for proposals will open soon: from 1 September until 1 December 2022.
Selected AI Projects 2021
The following 10 projects will receive funding of up to €400,000 each over a total funding period of two to three years.
- AI-4-XAS: Artificial Intelligence for X-ray Absorption Spectroscopy (HZB + Hereon)
The theoretical prediction of X-ray spectra for very large molecules is currently beyond the capabilities of the most powerful high-performance compute clusters. The core of this project is to utilize molecular graphs in combination with the probabilistic analysis of molecular motifs for a supervised machine learning prediction trained on data obtained for smaller building blocks of the huge molecules of interest.
Contact: Annika Bande, annika.bandenoSp@firstname.lastname@example.org (HZB)
- ALEGRA: Active Learning-enabled Generation of (Patho-)Physiological Lung Architectures for Pulmonary Medicine (DKFZ + HMGU)
ALEGRA aims to discover new insights into the morphology, physiology, and functionality of mammalian lungs. Therefore, quantitative parameters will be extracted from light sheet fluorescence microscopy images featuring comprehensive annotations of unprecedented detail for structures of interest such as hollow airways, blood vessels, as well as alveoli. These annotations are made possible by a novel active learning approach for semantic segmentation, nnActive, which will be developed as part of this project.
Contact: Paul Jäger, p.jaegernoSp@email@example.com (DKFZ)
- COMPUTING: COnnecting Membrane Pores and prodUction parameTers via machine learnING (Hereon + DKFZ)
Isoporous block-copolymer membranes play a crucial role in the separation of molecules from liquids and can be used, for example, for the purification of drinking water. Despite recent efforts, the production process of these membranes is still not understood, currently requiring multiple trial-and-error experiments to produce the desired output characteristics for each new batch of raw materials. This project aims at streamlining the manufacturing process and ultimately unlock the production of designer membranes by developing methods that predict the required production parameters given some desired morphology and raw material characteristics.
Contact: Martin Held, martin.heldnoSp@firstname.lastname@example.org (Hereon)
- DIADEM: Diabetes Detection from Histopathologic Images of Human Pancreas (HMGU + DKFZ)
Type 2 diabetes is a chronic, often debilitating disease, which is essentially determined by a complex dysfunction of pancreatic islets resulting in reduced insulin secretion. In this project, we aim to recognize patterns associated with islet dysfunction by leveraging deep learning in a unique set of histopathologic images with clinical and laboratory data from patients with and without type 2 diabetes. While key challenges of the project involve gigapixel-sized images with different histologic stainings and varying sample quality, explainability of the models will advance our current understanding of diabetes.
Contact: Robert Wagner, robert.wagnernoSp@email@example.com (HMGU)
- NACHMO: Neural Network-based Atmospheric Chemistry Module for Weather and Climate Models (Hereon + KIT)
Realistic atmospheric chemistry is crucial for simulating and predicting climate, weather and air quality, but is currently computationally infeasible. We will design and train neural networks to accelerate simulations of atmospheric chemistry, using specialized architectures and loss functions to enforce conservation laws, maintain accuracy and ensure long-term numerical stability. Our ultimate aim is to provide a fast, accurate chemistry module for weather, climate, and Earth system simulations beyond previous computational limits.
Contact: David Greenberg, david.greenbergnoSp@firstname.lastname@example.org (Hereon)
- Opt4Bio: principled optimization of structured deep learning models for multimodal biological data integration (HMGU + CISPA)
We will develop new efficient and theoretically sound optimization methods to train structured deep learning models for multimodal data integration. Such models are of particular interest in biology and health applications where datasets often consist of a mixture of structured and unstructured data (e.g., tabular data from high-throughput sequencing, and medical imaging data, respectively). Our focus on biological and health data is intended to facilitate the transfer of these techniques to clinically relevant applications, including fast biomarker detection in microbiome data.
Contact: Christian Müller, cmuellernoSp@email@example.com (HMGU)
- RESEAD: Robust Environmental Sensor data using Explainable data-driven Anomaly Detection (KIT + UFZ)
The increasing amount of sensor data requires robust automated quality control (QC). Within the project RESEAD we will go a significant step beyond existing QC methods by leveraging the full spatiotemporal information contained in the data of large distributed environmental sensor networks. We will develop a ready-to-use software pipeline consisting of a dense embedding method for sparse spatially distributed sensor data, a GAN-based data imputation, and a module to make the results of our DeepLearning-QC-pipeline explainable.
Contact: Christian Chwala, christian.chwalanoSp@firstname.lastname@example.org (KIT)
- SedimentAI: Automated particle detection in Sediments using imaging flow-cytometry and Artificial Intelligence (UFZ + AWI)
Past climate reconstruction based on paleobotanical records of sediments is limited in throughput due to manual counts of experts. In SedimentAI we want to test the innovative combination of multispectral imaging flow cytometry and a semi-supervised learning approach for this challenge. The project will substantially advance the field of paleobotany and would furthermore provide a deeper understanding of past ecosystem dynamics with respect to climate change.
Contact: Susanne Dunker, susanne.dunkernoSp@email@example.com (UFZ)
- SURF: (Semi/Un)supervised machine learning flood damage assessment (DLR + GFZ)
In the last decade, flood events have left 1 billion people without a home. To improve the early response of crisis management, SURF aims to develop an approach to rapidly assess building damage after flood events. Particularly, we propose to explore three machine learning-based approaches to provide a near real-time solution that benefits from existing global urban mapping data and information collected in the early stage of the flood.
Contact: Andrés Camero Unzueta, andres.camerounzuetanoSp@firstname.lastname@example.org (DLR)
- UNITY: UNcertainty and explainabilITY for unsupervised deep learning (MDC + DZNE)
The old adage that AI is a “black box” is quickly eroding: Algorithms from "explainable AI" can identify the most important input features, and neural networks can provide uncertainty estimates about predictions and their interpretations. UNITY will push the development of such uncertainty and interpretability approaches for the example application of unsupervised integration of multimodal, sparse, high dimensional genomics data, to increase our molecular understanding of disease trajectories.
Contact: Uwe Ohler, uwe.ohlernoSp@email@example.com (MDC)