Many young people experience adversity while growing up, such as bullying, neglect, or abuse, which increases the risk of mental health problems later in life. Yet, some remain resilient. Resilience refers to the ability to maintain or regain mental health and well-being despite these challenges. In the THRIVE study, we investigate why some individuals are more resilient than others.

We follow young adults (18–24 years) over time and combine questionnaires, behavioral tasks, hormonal measures, and brain imaging to understand how social interactions, cognitive processes, and stress responses relate to mental health. For example, we study how individuals process emotions, learn from feedback, and respond to social stress.

From a data perspective, this project involves integrating complex, multimodal, and longitudinal datasets. Imaging analyses are performed on the ALICE high-performance computing cluster, while other sensitive data (e.g., behavioral and clinical measures) are stored and processed in separate, secure environments. This setup reflects both computational demands and strict data governance requirements. The project also involves methodological challenges, such as reliably de-identifying anatomical MRI scans to prevent participant recognition while preserving data quality for analysis.

By uncovering how social support and cognitive functioning contribute to resilience, the THRIVE study aims to inform interventions that strengthen mental health in young people.

Scheduling jobs is the core of how we interface with our clusters. But how do we schedule jobs optimally? How do we ensure we're not wasting resources unnecessarily? This talk will look at some prior art for optimising cluster scheduling as well as the author's recent work towards improving job runtime and memory requirement with linear models based on historical job performance data, in the context of WDL pipelines executed within Klinische Genetica.

I will describe image pipelines I ran on ALICE for a dissertation on colonial Korean print
culture. Two pipelines sit at the centre. One uses a YOLO detector to extract 600,000
advertisement cutouts from digitised Korean newspapers, then embeds them with a DINOv3
ViT-B backbone fine-tuned with rank-8 DoRA adapters. The other starts from 57,583 page
images of four colonial printshops and produces several million attention-weighted patches
through a multi-instance-learning model trained from scratch on ALICE.

The hardest questions were not about input rates but about storage. Do we keep raw image
files on disk, or move everything into a database? Each pipeline started from raw files and
ended with a database of cutouts and embeddings that downstream code could query without
touching the scans again. For the advertisement pipeline, I ran an entire round of experiments
on 200-pixel thumbnails before committing ALICE time to full-resolution crops.

After training, the pipelines run YOLO, fine-tuned DINOv3, and the MIL model over the
full corpus to populate the database. The downstream clustering first used scikit-learn UMAP
and HDBSCAN on CPU nodes, which collapsed past the million-patch mark. Swapping in
NVIDIA cuML on a single GPU required essentially no code changes and turned days into
hours, cheap enough to iterate on rather than run once.

I will close with a few small best practices I have picked up along the way.

In this talk, we introduce slurm-board, an interactive dashboard designed to simplify the configuration, monitoring and submitting of Slurm jobs. Users often find it difficult to configure a job due to complex settings and limited knowledge of available resources, resulting in inefficient cluster usage and longer queue times.

Slurm-board addresses these issues by providing a user-friendly SBATCH-generator that allows users to easily configure jobs. Combined with real-time resource availability information, it enables users to efficiently configure their jobs