We would like to thank all applicants for computation time within the Special Call: GPU Testing and Benchmarking LUMI-G.

Within this call, a total of 7 projects were submitted, among which 47 400 node hours on the accelerated parts of the LUMI supercomputer were distributed. 

---

The Allocation Commission decided on the allocations within the Special Call: GPU Testing and Benchmarking LUMI-G as follows:

 

Researcher: Petr Kouba         
Project: Machine learning for dynamics-aware protein design
Allocation: 8 000 nodehours
Abstract: The problem of protein design is typically understood as a problem of finding a proper sequence of amino acids defining the desired protein 3D structure, dynamics, and consequently its function. The machine learning tools tackling this problem are currently rapidly evolving. For example, tools like ProteinMPNN [1] are significantly pushing the state-of-the-art for tasks of designing a protein sequence given its 3D structure. In our project, we strive to enhance the applicability of such an approach to the novel task of designing a protein sequence given both the required 3D structure and the required dynamics. In particular, we aspire to develop a framework in which structures of multiple conformations of a single protein, potentially with requirements on flexibility of particular regions, would be provided on the input and an appropriate protein sequence would be obtained on the output.

 

Researcher: Jakub Kopko     
Project: Deep learning-based analysis of the dimerization process of Apolipoprotein-E
Allocation: 3 400 nodehours
Abstract: This study focuses on investigating new hypotheses regarding the impact of Apolipoprotein-E (APOE) dimerization on Alzheimer's disease (AD). Mutations in the APOE gene have been identified as a significant genetic risk factor for the development and progression of AD. This study aims to understand the potential influence of APOE aggregation on the development of the disease. Molecular dynamics (MD) simulations allow the analysis of the physical motion of biomolecules. The generated data are sequences of frames (hundreds of thousands of frames per simulated system) captured at a predefined time step. Each frame consists of the positions of all the atoms (from hundreds to tens of thousands) of the biomolecules. The objective of this project is to investigate the APOE dynamics by existing machine learning tools, in particular the VAMPnet neural network [1], and to develop their extensions utilizing graph neural network architecture, similar to [2]. The results and complexity of all the methods will be thoroughly analyzed. By elucidating the dynamic properties of APOE and its aggregation patterns, this work may contribute to the identification of novel strategies for mitigating the effects of APOE-related pathology in AD treatment and prevention.

 

Researcher: Anton Bushuiev
Project: Machine learning for the design of protein-protein interactions
Allocation: 8 000 nodehours
Abstract: Stroke is a leading cause of death and disability worldwide, resulting in one of the heaviest socioeconomic burdens of any disease kind. In this project, we will apply state-of-the-art machine learning methods to design a next-generation thrombolytic staphylokinase. The computational experiments will involve the engineering and optimization of specific parts of the staphylokinase protein using recent machine-learning methods RFdiffusion and dl_binder_design. In addition, we will improve our novel deep-learning method PPIformer, which has demonstrated strong potential in advancing the current state of the art in modeling protein-protein interactions. We will utilize PPIformer to enhance the thrombolytic activity of staphylokinase, as it is directly influenced by the interactions between staphylokinase and other proteins.

 

Researcher: Michal Vavrečka
Project: Robotic manipulation tasks trained with multi-GPU distributed reinforcement learning
Allocation: 8 000 nodehours
Abstract: Recent progress in robotic simulation allows to train reinforcement learning at the unprecedented speed of 100,000 frames per second [1] that is 10 years of human experience in 1 day. On the other hand, these simulated environments have simplified physics resulting in a drop from 97 percent accuracy during training to 30 percent during testing in environments with rich physics [1]. We plan to adopt fast decentralized algorithms [2] from the simplified environments [1] and adopt them into our more physically complex simulators [3] and test their accuracy in robotic manipulation tasks. Successful training will result in better sim2real transfer.

 

Researcher: Vladimir Petrik
Project: Object 6D pose and camera focal length estimation for objects unseen during training
Allocation: 8 000 nodehours
Abstract: Estimating the pose of the object in the world is a crucial capability that allows a robot to build object-centric representations, for example, for planning manipulation motion [1]. High accuracy of the 6D poses estimation from images has been achieved recently by the render-and-compare methodology in CosyPose [2] for calibrated cameras and in FocalPose [3] for uncalibrated cameras. However, both CosyPose and FocalPose require costly (a few hours on 40 GPUs, a few days on 8 GPUs) retraining for every new object, which complicates and slows down their applications in practice. For calibrated cameras, MegaPose [4] addressed this issue by training on a big synthetic dataset that generalizes well to novel objects, i.e., objects unseen during training. Our goal is to extend MegaPose to an uncalibrated camera setup, which would allow us to estimate the poses of novel objects in the wild, for example, from YouTube videos where camera parameters are unknown. Such a capability can be directly used for learning or planning robotic manipulation skills from online instructional videos or images without metadata, and would lead to explainable object-centric representation without costly retraining for every new object.

 

Researcher: Jiri Brabec
Project: GPU-accelerated masivelly parallel DMRG implementation
Allocation: 4 000 nodehours
Abstract: The density matrix renormalization group (DMRG)-based methods are undoubtedly the most popular choice for a proper description of the strongly correlated systems, where static and dynamical correlation effects are in play. Recently, we have developed the masivelly parallel DMRG code (MOLMPS) [molmps], which exploits operator and symmetry sector parallelisms. In this project, we will extend the parallel scheme of the computationally most demanding parts by the GPU acceleration, which should significantly speedup the execution, extend the applicability, and increase the effectiveness of the program.

 

Researcher: Radim Špetlík
Project: Applications of Denoising Diffusion Probabilistic Models
Allocation: 8 000 nodehours
Abstract: Machine learning is crucial for computer vision. It enables automatic feature extraction, accurate object detection and recognition, precise semantic segmentation, image captioning, video analysis, and transfer learning. It empowers computers to understand visual data like humans, leading to applications in autonomous driving, surveillance, medical imaging, augmented reality, and more. Denoising diffusion probabilistic models (DDPMs) are generative models that use deep neural networks to model the joint distribution of clean images and noisy observations. Through a diffusion process, the model gradually transforms noisy images into clean ones. DDPMs capture complex dependencies and structures, leading to high-quality denoising results. Due to its universality, DDPM may be applied to a variety of problems including all mentioned above. We are primarily interested in two areas of DDPM applications: (i) In image style-transfer, DDPMs can be utilized to transform images from one style to another while preserving the content. This enables applications such as artistic rendering, photo editing, and visual effects. (ii) Temporal super-resolution involves enhancing the temporal resolution of a video sequence, creating intermediate frames between existing frames to generate a smoother and more detailed video. This has applications in video restoration, slow-motion video generation, and video compression.