VSR Seminar with Two Talks

1st Talk: Foundational deep learning models for robust cell instance segmentation

Speaker:
Eric Upschulte, Institute of Neurosciences and Medicine, INM-1

Abstract:

Understanding and quantifying brain structure at cellular resolution requires the analysis of massive microscopy datasets, from which cell bodies, tissue architecture, and other biologically meaningful structures must be extracted and quantified reliably. State-of-the-art solutions for this task are based on deep learning, but achieving the robustness required for real-world large-scale analysis demands both diverse training data and substantial HPC resources. We develop scalable deep learning methods for large-scale cytoarchitectonic analysis of human brain tissue at cellular resolution. The project targets two core problems: universal instance segmentation across heterogeneous staining protocols, imaging domains, and object classes, and generative inpainting of missing or damaged tissue to enable consistent 3D reconstruction of cell distributions. On the discriminative side, we extend our generalist training pipeline through continuous large-scale dataset integration (SciDa) and investigate contour-based formulations that are well suited to biomedical imagery with thousands of instances per crop, building on our established Contour Proposal Networks (CPN) and advancing newer Contour Detection Transformers (CDETR) and promptable Contour Anything Models (CAM) within the CellDetection framework. Because these models scale efficiently in both training and inference and are trained on a broad collection of datasets, they are also applicable to a wide range of scientific imaging segmentation tasks beyond the present use case. On the generative side, we build on our previously published diffusion-based inpainting framework and extend it toward multi-section reconstruction, faster consistency-style models, automated artifact detection, and multi-brain generalization. Recent large-scale runs on JURECA-DC, JUWELS Booster, JEDI, and JUPITER demonstrate strong multi-GPU scalability and provide the technical basis for larger model and data scaling experiments. The project is supported by annual allocations of 13-14 million core-hours and combines smaller runs for model development and hyperparameter tuning with larger-scale runs for validated models and configurations, as well as high-throughput inference workloads.

2nd Talk: High Definition Simulation of Packed Bed Chromatography

Speaker:
Xiang Xu, Institute of Bio- and Geosciences, IBG-1

Abstract:

Packed-bed liquid chromatography is widely utilized in biotechnology and related industries as a critical unit operation in downstream processing, i.e., the separation and purification of valuable molecules from unwanted side products and other impurities. While reduced-order models (ROMs) enable efficient simulations, they rely on extensive experimental calibration and lack mechanistic insights into how specific morphological features influence local hydrodynamics and separation performance. These features include particle size distribution, particle geometry, local packing irregularities, and wall effects. High-definition (HD) modeling in this project addresses these limitations by resolving full spatial details. The coupled multi-physics, multi-domain chromatography problem is simulated using the highly parallel solver XNS with a stabilized space–time finite element method on the JURECA supercomputer at Forschungszentrum Jülich. The computational model setup and parallelization strategies will be presented.

This study first examines confined cylindrical packings with up to 10,000 particles, representative of narrow columns, enabling a detailed understanding of the chromatographic separation process. The impact of particle size distribution on velocity and concentration profiles, as well as breakthrough curves, is systematically characterized [1]. Analysis of the flow field reveals hotspots, where the maximum axial velocity reaches approximately an order of magnitude higher than the average velocity predicted by the homogeneity assumption commonly adopted in ROMs. Visualizations of mass transfer simulations further show non-concentric particle loading and early breakthrough near the column wall in both monodisperse and polydisperse packings. This early breakthrough is attributable to pronounced wall effects caused by increased porosity in the near-wall region, which distorts local flow behavior and constrains the applicability of confined geometries to larger-scale systems.

Building on these findings, HD chromatography simulations are further extended from laterally confined to unconfined compartments through a validated workflow encompassing periodic packing generation, meshing, partitioning, CFD simulation, and post-processing. Unconfined HD modeling provides a representative unit cell that enables efficient scale-up investigations across laboratory and industrial scales, where wall effects become negligible due to the sufficiently large column-to-particle diameter ratio. The resulting intra-column data are used to calibrate axial dispersion and film diffusion parameters in the corresponding reduced-order model within the open-source bioprocess software CADET. The calibrated ROMs can efficiently approximate a wide range of packing configurations and be used for rapid process design and optimization [2]. The quantitative impact of particle size distribution and wall effects on these fundamental parameters will be discussed in the presentation.

[1] Rao J. S., Püttmann A., Khirevich S., Tallarek U., Geuzaine C., Behr M., & von Lieres E., High-definition simulation of packed-bed liquid chromatography, Computers & Chemical Engineering, Vol. 178, 108355, 2023.

[2] Rao J. S., Leweke S., Breuer J. M., Menzel S., Behr M., & von Lieres E., Two-dimensional general rate model with particle size distribution in CADET calibrated with high-definition CFD simulated intra-column data, Separation and Purification Technology, 134409, 2025.