MDSI Kickstarter Seed-Funds

Active Learning of Graph Neural Network Potentials from Quantum and Experimental Data (Active Learning Q&E) develops data-efficient graph neural network potentials using quantum (DFT) and experimental data. Active learning is used to select high-uncertainty configurations for labeling. It focuses on efficient uncertainty quantification methods such as deep-kernel learning and evidential deep learning. The approaches are evaluated on molecular dynamics data for accuracy and out-of-distribution detection.

Advancing spatiotemporal pattern analysis using top-level sports data (ASPAS) analyzes large-scale human spatiotemporal movement data using machine learning. It focuses on detecting patterns and learning compact trajectory representations via semi-supervised learning. The methods are developed and tested on soccer tracking data. The learned representations support clustering and prediction of movement patterns.

Algebraic Topology Preservation Learning for Image Segmentation (ATPL4IS) improves medical image segmentation by preserving topological structures such as vessels or membranes. It integrates persistent homology into deep learning through a novel topological loss function. This enforces consistency between predicted and ground truth segmentations. The approach improves segmentation quality by maintaining geometric and topological relationships.

ClimVine: Drought and Late-Frost Risk in a Changing Climate studies the joint risk of drought and late frost under climate change. It introduces a statistical framework using a Y-vine copula model to capture their dependency. The methods are applied to climate data for Bavaria. The results support improved forest risk assessment and ecosystem management.

Deep Learning for De Novo Peptide Sequencing (DL4DNPS) develops deep learning methods for de novo peptide sequencing from mass spectrometry data. It introduces tools such as Koina and Oktoberfest to standardize and accelerate peptide analysis. Building on these, the Spectralis model improves sequence prediction and confidence scoring. The approach enhances scalability and accuracy in proteomics applications.

3D Human Motion Capture for Cooperative Construction Robotics (HuMoCap) develops a 3D human motion capture system for safe human–robot collaboration in construction. It combines visual-inertial SLAM with deep learning to estimate full-body pose under challenging conditions. The system is applied to cooperative tasks such as brick laying. It enables accurate tracking and supports more autonomous collaboration.

HybridD3: Data Insights for Whom? Hybrid Data-Driven Decision-Making in Educational Contexts studies hybrid human–AI decision-making in higher education. It compares how learners use data insights through AI tutors and human-driven interactions. Using the ARTEMIS platform, it analyzes learning outcomes based on real-time student data. The goal is to design effective data-informed learning systems.

Mathematics-driven environmental sensing (MES) develops mathematical methods for environmental sensing of urban greenhouse gas emissions. It uses sensor data and compressed sensing techniques to localize emission sources. The approach addresses challenges such as measurement coherence and uncertainty. The goal is to improve the accuracy of urban climate monitoring.

Machine Learning for Medicine: Accelerating Metadynamics of Supramolecular Host-Guest Complexes for Therapy and Imaging (MiAMI) develops machine learning–based methods to study supramolecular host–guest systems for therapy and imaging. It combines quantum chemistry, force fields, and ML interatomic potentials to model structural dynamics. The approach evaluates how different modeling choices affect stability and flexibility.

Mobilität.Leben.daTUM (M.L.daTUM) analyzes the impact of Germany’s €9 public transport ticket using survey, app-based mobility tracking, and traffic data. It combines individual behavior and city-wide mobility patterns. The dataset supports research in mobility, economics, and policy. The goal is to identify promising directions for further studies.

Multi-Modal Foundation Models for Particle Physics under Model Misspecification (MMFM) develops scalable multi-modal foundation models for particle physics. It combines self-supervised learning and alignment methods to handle domain shifts between simulation and real data. The models support tasks such as classification and particle identification. The goal is to build general-purpose models for particle physics.

Netrium: Precise and Fast Model Prediction with Machine Learning develops a machine learning model to accelerate neutrino mass analysis in the KATRIN experiment. The model predicts the tritium beta-decay spectrum with high precision from physical parameters. It significantly reduces computation time while maintaining accuracy. This enables faster and more efficient data analysis. 

Fast and Accurate 3D Image Segmentation by Matching Topological Features (Segmatch3D) develops fast, topology-preserving methods for 3D medical image segmentation. It extends persistent homology–based matching to volumetric data. A high-performance implementation enables efficient training on large datasets. The approach supports applications such as vessel segmentation and cell tracking.

Seismic Safety studies the impact of geothermal-induced micro-seismic events on buildings. It uses sensor data and machine learning to model and predict structural responses. Surrogate models are developed to quantify uncertainty and improve predictions. The goal is to enable vibration forecasting and enhance safety in geothermal systems.

Synthetic Benchmark Datasets for Finance (SyBenDaFin) develops synthetic benchmark datasets for finance to support machine learning research. It addresses data scarcity and privacy constraints in real financial data. Using simulation and generative methods, it creates accessible reference datasets with quantified uncertainty. The goal is to establish benchmarks for evaluating financial ML models.

VineGP: Learning Predictive Vine Copula Models for Complex Plant Traits develops predictive vine copula models for genomic prediction of plant traits. It captures complex, nonlinear dependencies between genetic and environmental factors. The approach improves prediction accuracy and helps identify influential SNPs. It provides new tools for plant breeding.

Explainable AI to Understand the Impact of Life Events on Travel Behavior Change (XAITraB) studies how life events influence travel behavior using explainable AI. It analyzes changes such as trip length, activity patterns, and transport mode. Using longitudinal travel data, it captures both individual and combined effects of life events. The goal is to better explain travel behavior compared to traditional models.