Student Posters Repository

In recent decades, the analysis and characterization of solid-solid interfaces have become increasingly important. These interfaces are critical for electrochemical and diffusion processes that underpin the development of new energy and information storage devices, such as batteries and chips. Historically, experimental investigation of these interfaces has posed significant challenges, often necessitating support from theoretical simulations. Such simulations have typically relied on density functional theory or classical force fields. However, these methods have been constrained by extensive computational demands or by the low accuracy of the resulting physical quantities.

The advent of machine learning within theoretical chemistry now offers a promising avenue to overcome these limitations, enabling the simulation of large interfacial structures which was previously infeasible. In this contribution, I outline a strategy for developing a neural network force field aimed at simulating key processes at the interfaces of SiC and Ti. 

Tensor train (TT) decomposition is a method for approximating and analysing tensors. TT-SVD, the most commonly used TT decomposition algorithm, computes the TT-format of a tensor in a sequential manner by alternately reshaping and compressing the tensor. For large tensors, this requires a large amount of computation time and memory. In this paper, we propose a distributed parallel algorithm, PTTD, to perform TT decomposition, which distributes parts of the tensor to all processes, decomposes it in parallel using TT-SVD, and merges the results to obtain the TT-format of the original tensor. Rounding is applied to reduce the size of the merged TT-formats. The algorithm is deterministic, which means that approximation error is controllable and there is no need to know the TT-ranks of the tensor in advance. Experimental results show that PTTD achieves an average speedup of 5384× using 8192 cores, and that the approximation error decreases as the number of cores increases, at the cost of slowly growing TT-ranks.

The utilization of generative Artificial Intelligence (AI) for molecular design represents a cutting-edge frontier in computational chemistry, offering unprecedented potential for innovation across drug discovery and broader chemical research domains. Our research aims to extend the capabilities of generative models to produce 100% chemically valid molecules, covering organic chemical research beyond traditional drug design. Despite significant advancements in previous research, a gap remains in achieving universally valid molecule generation and developing models adaptable to diverse datasets without being constrained by their specific distributions.

To address these challenges, we introduce the MolGen-Transformer, which is based on the Simplified Self-Referencing Embedded Strings (SELFIES) representation and trained on a meta dataset encompassing a vast array of organic molecules to ensure the validity and diversity of generated molecules. We also developed a multi-layer transfer learning hierarchy to enhance model adaptability and efficiency across different datasets and introduced the Neighboring Search and Optimistic molecule evolution algorithms for tailored molecular design.

We demonstrated the MolGen-Transformer's ability to achieve 100% reconstruction accuracy across extensive molecular datasets and showcased its utility in generating diverse molecular structures. The multi-layer transfer learning framework highlighted the model's predictive accuracy across varied datasets. Additionally, the Neighboring Search algorithm facilitated the generation of molecules closely matching target structures. Furthermore, the Optimistic Molecules Evolution algorithm advanced molecular design by enhancing properties using Multi-Fidelity Models, validated through Density-Functional Theory (DFT) simulations, showcasing a holistic approach to precise and efficient molecular innovation. Overall, our work paves the way for a new era in molecular design, offering versatile, efficient, and broadly applicable tools for the chemistry research community.

Pre-training large-scale vision-language foundation models is crucial for enabling robust and native support for multimodal applications. Large language model pretraining has shown significant advancements in understanding and generating human-like text. Similarly, training vision-language models from scratch is essential to achieve seamless integration support for both visual and textual data. In this work, we address the challenges of pre-training vision-language models by developing a practical and efficient training codebase using auto regressive loss. We accelerate training processes via 3D parallelism and leverage the WebDataset format to optimize IO speed. This combination significantly improves training efficiency and scalability, laying the foundation for robust vision-language models.

Superficial white matter (SWM) has been less studied than long-range connections despite being of interest to clinical research, and few tractography parcellation methods have been adapted to SWM. Here, we use diffusion MRI data and an in-house efficient geometric-based parcellation method (GeoLab) that allows high-performance segmentation of hundreds of short white matter bundles from a subject. This method has been designed for the SWM atlas of EBRAINS European infrastructure, which is composed of 657 bundles and it has been applied to different atlases. The atlas projection relies on the precomputed statistics of six bundle-specific geometrical properties of atlas streamlines. In the spirit of RecoBundles, a global and local streamline-based registration (SBR) is used to align the subject to the atlas space. Then, the streamlines are labeled taking into account the six geometrical parameters describing the similarity to the streamlines in the model bundle. After the bundles were extracted for all participants, we computed 5 different diffusion measures for each bunde in order to study the ageing trajectories of SWM.

Understanding the rarity of circumbinary disks (CBDs) compared to single-star disks involves investigating fundamental differences between these systems. Observational data suggest that CBDs are less common than their single-star counterparts, prompting inquiries into the underlying mechanisms driving this discrepancy. Computational simulations, notably employing the DISCO code, facilitate modeling the evolution of circumstellar and circumbinary disks, shedding light on their behavior over time. Leveraging DISCO's capabilities, scientists discern disparities in depletion rates and longevity between single-star and binary systems. This exploration aims to unravel the factors contributing to CBD rarity compared to disks around single stars. Understanding these differences is crucial for elucidating the complexities of star formation and binary companions' roles in shaping circumstellar and circumbinary environments. By combining observational data analysis, theoretical modeling, and computational simulations, researchers advance our understanding of CBDs and their scarcity relative to single-star disks. We observed that varying the disk size while keeping all other conditions constant, the disk exhibited a tendency to approach an infinite state, revealing a new power law in our simulated disk observations. Specifically, the disk lifetime, denoted as tau, seemed to extend further as the disk size increased. This observed extension in disk lifetime is described in detail in the discussion and methods sections of this paper.

Metacognitive monitoring involves evaluating one's own thought processes and state of knowledge. Pupils with higher educational achievement consistently demonstrate good skills in self-monitoring their past performance. In recent efforts to design intelligent tutoring system (ITS) for educational purposes, scaffolds for the metacognitive monitoring process have been implemented to foster higher learning outcome. Our work is subject in training deep-learning networks to interpret facial expressions while performing metacognitive monitoring process and design the AI agent to interact with learner based on the estimated metacognitive monitoring performance.  

So far, we built the video dataset (Affect2Metacognition) to train and test state-of-the-art neural networks for metacognitive monitoring performance identification. In addition, we already proposed a strategy for scaffolding pupils' metacognitive monitoring process (under review), which is based on identifying their metacognitive monitoring performance through facial expressions. In the long term of our research, we will analyse the impact of this strategy on pupils' learning outcomes.

Atmospheric new particle formation (NPF) leads to the formation of around half of the cloud condensation nuclei needed to form the seeds of cloud droplets. Therefore, the participation of primarily anthropogenic species such as ammonia in NPF may influence the radiative forcing of climate. However, most climate models do not account for how ammonia helps sulfuric acid to form new aerosols. Here we incorporate a ternary NPF parameterization into the UK Met Office Unified Model (UM) using results from the CERN CLOUD chamber. The parameterization also includes the influence of ions produced from radon and cosmic rays. We test the parameterization in a one-way nested UM configuration, where a regional simulation with a 3km horizontal grid resolution over the Colorado Front Range region (centered at 40.0, -105) runs inside a global simulation(N96). The model simulates the 2014 Summer period (July 20th, 2014 to August 10th, 2014) when we are able to evaluate both aerosol and precursor concentrations using measurement data from the DISCOVER-AQ and FRAPPE field campaigns.

This research paper explores the development of advanced methods for the detection and management of sporadic rogue processes in High-Performance Computing (HPC) systems. Rogue processes, often manifesting as unauthorized or unexpected resource consumption, pose significant challenges to the efficiency, security, and reliability of HPC environments. This study introduces a novel machine learning-based approach for real-time anomaly detection, coupled with an adaptive resource management system. The proposed solution aims not only to detect rogue processes more accurately but also to respond proactively to mitigate their impact. Evaluations conducted on several HPC environments demonstrate the effectiveness of this approach in improving system performance and security. This work contributes to the broader field of HPC system management and offers a scalable solution adaptable to various computing environments. Introduction High-Performance Computing systems are integral to modern computational research and data processing. Their efficiency and security, however, are frequently compromised by sporadic rogue processes - unauthorized or unexpected tasks that consume disproportionate system resources., Definition In this context, a "rogue process" is defined as any process that operates outside established norms for resource usage in an HPC environment. These processes may be unintentionally created due to programming errors, misconfiguration, or system anomalies and can significantly impact system performance. Traditional methods for detecting and managing rogue processes in HPC systems rely on static threshold-based monitoring, lacking the flexibility and accuracy needed for dynamic HPC environments. Consequently, these methods are often reactive, addressing issues only after they have negatively impacted system performance. This research aims to develop an intelligent, machine learning-driven approach for the real-time detection of rogue processes in HPC systems. The study also explores the integration of this detection system with an adaptive resource management framework to proactively respond to and mitigate the impact of these processes. This study encompasses the development and validation of the proposed solution within various HPC environments. Its significance lies in its potential to enhance the reliability and security of HPC systems, thereby supporting a broad spectrum of scientific and industrial applications.