Accelerating AI for subsea microscopy in ocean monitoring and aquaculture

The organisation

FP AQUA is a Danish AquaTech startup specialising in continuous underwater microscopy and using AI (ResNet / YOLO / Image Recognition / SAM / Autoencoders) to automatically train, classify and identify microscopic lifeforms in the ocean. They are at the forefront of transforming aquaculture with innovative technology that addresses critical challenges for industry and biodiversity. Their technology focus is dual-purpose:

• To support aquaculture operators, for example, in salmon farming, by providing early warnings of emerging threats, offering real-time monitoring solutions to combat parasites such as sea lice, micro-jellyfish and other harmful zooplankton.
• To monitor the biodiversity of the oceans.

The challenge

FP AQUA’s core technological challenge is the large-scale processing and analysis of ultra-high-resolution underwater microscopy data. With three active sensors operating continuously at different locations in Norway, the company currently collects approximately 6,000 images per day, each with a resolution of up to 64 megapixels. Even after extensive filtering, the resulting data volume exceeds the processing capacity of the existing hardware infrastructure.

A critical bottleneck lies in image segmentation and model training. Due to GPU memory limitations, full-resolution images must first be subdivided before analysis. As a result, processing a single day of collected data required nearly six days of computation, preventing near real-time analysis and limiting the scalability of the overall workflow. Based on current estimates, at least six times more computational capacity would be required to match the daily data ingestion rate, while a substantially larger increase would be needed to support the complete processing pipeline at scale.

In parallel, FP AQUA aims to advance its AI-based analysis capabilities. The marine environment contains an extremely large diversity of planktonic organisms, including zooplankton, phytoplankton, and microscopic jellyfish, many of which exhibit subtle morphological and chromatic differences. To address this challenge, the company seeks to develop and train advanced autoencoder models capable of automatically clustering detected organisms into biologically meaningful species groups based on visual similarities.

Furthermore, FP AQUA aims to significantly accelerate the training of its existing computer vision models, including YOLO-based object detection systems, while maintaining or increasing image resolution. Current training cycles require approximately 30 hours even at reduced image resolution, creating a constraint on model development, experimentation, and deployment.

The solution

EuroCC Denmark supported FP AQUA in accessing EuroHPC resources to accelerate two key AI workflows: autoclustering with autoencoders and object detection with YOLO. These models now run on the MeluXina GPU AI Factory, which offers the CUDA-compatible Nvidia GPUs they rely on. The EuroCC Denmark team facilitated the transition from local infrastructure to HPC, ensuring that image segmentation and classification tasks could scale efficiently.

Business impact

With access to European supercomputers, FP AQUA can now process and analyse millions of images in a fraction of the time. This leap in performance enables faster detection of emerging threats in aquaculture and more effective monitoring of ocean biodiversity, all without the burden of in-house server maintenance or debugging.

Benefits

• Speed is the largest benefit as compared to the single AMD Threadripper and single Nvidia 3090 we are running on now.
• Our own server can run our other processes, when having some of our heaviest running on the AI Factory.
• Maintenance is done by the provider, so we don’t need to use all our time debugging our server.

All use cases