Overview
The AI for Science Group at TH Köln develops machine learning methods infused with scientific knowledge to enable accurate simulations, parameter inference, and enhanced interpretability for applications in science and industry. Our research spans semiconductor metrology, quantum computing, biology, and ML engineering.
Metrology + Machine Learning
Overview
We develop object-centric learning and physics-informed generative models for semiconductor metrology, enabling layer separation and process parameter estimation from complex imaging data.
Key Topics
- Object-Centric Learning (OrthoRF): Orthogonal rotating features for unsupervised decomposition of overlapping structures in SEM images
- Latent Space Alignment: Cross-device inference via aligned autoencoder latent spaces
- Process Parameter Estimation: Autoencoders for extracting physical parameters from multi-sensor metrology data
- Physics-Informed Generative Models: VAE variants with domain-specific priors and equivariance properties
Recent Highlights
- OrthoRF accepted at ICLR 2026: Competitive binding + orthogonality loss for discrete object representations without post-hoc clustering
Collaborators
- ASML Eindhoven
- University of Amsterdam
- RWTH Aachen University
- University of Cologne
- Aixtron
Quantum Computing + Machine Learning
Overview
We use physics-informed generative AI for quantum hardware calibration and control, building digital twins that combine simulation and experimental data with structured latent spaces.
Key Topics
- Digital Twins for Quantum Devices: Physics-informed autoencoders combining simulation and experimental data
- Electron Shuttling Optimization: VAE-based reconstruction and optimization of single-electron transport in quantum processors
- Qubit Characterization: Deep learning for data-driven extraction of qubit parameters
- Charge-Stability Diagram Analysis: Autoencoders for extracting control parameters from charge stability diagrams
- Valley-Splitting Reconstruction: Inverse problem solving for spatial landscape reconstruction during shuttling
Funded Projects
- KIMO-QUANT (~590k€, HAW-ForschungsAkzente, 2027–2030): GenAI for quantum hardware optimization with ML4Q partners
Collaborators
- ML4Q Cluster (RWTH Aachen, Uni Koeln, FZ Juelich)
Biology + Machine Learning
Overview
We develop AI models for non-destructive plant biomass estimation, combining computer vision with environmental sensor data for climate adaptation research.
Key Topics
- Biomass Estimation: CNN regression for non-destructive measurement of plant biomass from images
- Multimodal Models: Incorporating metadata (age, weather, sensor data) alongside image inputs
- Synthetic Data Generation: Training on synthetic plant images and using multimodal LLMs for data preprocessing
- Growth Model Priors: Embedding biological domain knowledge into VAEs for improved predictions
Results
- Achieved ~20% MAPE for Arabidopsis thaliana biomass estimation (highest known accuracy in given weight range)
- Automated quality filtering for improved prediction reliability
Collaborators
- Bucher Lab (University of Cologne)
ML Engineering (MLOps)
Overview
We build reusable MLOps systems for scientific computing that enable reproducibility, collaboration, and easy deployment for research ML projects.
Key Topics
- MLOps for Science: Data versioning, experiment tracking (MLflow), pipeline orchestration (Dagster)
- Domain Knowledge Embedding: Auto-extracting knowledge from PDFs via LLMs and embedding it into ML models
- Low-Code/No-Code AI Builder: Tools for ML applications in domains like biomass or quantum data
- KI-Twins: AI twins of simulation frameworks
Key Insights
- Wrong domain knowledge degrades model performance; correct knowledge yields significant improvement
- Reusable MLOps pipelines reduce experiment overhead and improve reproducibility across projects
Industry Collaborations
Overview
We maintain active collaborations with industry partners on applied AI research.
Partners
- ASML: Object-centric learning for semiconductor metrology, LLM-assisted cost function design
- KI-CyberVision: Robot vision and AI cyber security for automotive manufacturing (Adient Metal Forming Technologies)
Current Student Projects
- MLOps Platform for Scientific Computing — Building a reusable MLOps infrastructure for research ML projects
- LLM-Guided Equation Solving — Using large language models to assist in solving systems of linear equations
- Quantum Dot Charge Diagram Analysis — Autoencoder-based extraction of quantum dot state parameters from charge stability diagrams
- Conveyor Simulation & Valley-Splitting Extraction — Simulation-based optimization and inverse problem solving for electron shuttling
- Synthetic Plant Image Generation — Training biomass models on synthetically generated plant images
- Multimodal LLM Data Preprocessing for Biomass — Using multimodal LLMs to pre-sort and filter biomass datasets
- AI-Assisted Oral Tests — AI voice-assistant for conducting oral test
- Training Gap Detection in ML Datasets — Identifying and addressing gaps in training data distributions
- Non-Destructive Biomass Estimation — CNN-based plant biomass prediction from image and sensor data
- Digital Twins of Simulation Frameworks — AI twins replicating the behavior of physics simulation tools
- AI-Powered Learning Platform — Intelligent tutoring and learning support platform
- Discrete Diffusion Models for Quantum Error Correction — Improving accuracy of quantum error correction decoders via discrete diffusion model
- EDA Tools for Quantum Computing — Electronic design automation pipelines for shuttling-based quantum computing architectures
- Leaf Overlap Estimation — Image-based detection and quantification of leaf overlap in plant canopies
Funding
Our research is supported by:
- BMFTR
- MWIKE
- European Commission
- EFRE
- TH Köln