ACM Posters – Student Research Competition
Designing Biomimetic Materials for Carbon Capture: Leveraging High-Performance Computing for Large-Scale Molecular Dynamics Simulations to Advance Sustainable Solutions
Sustainable carbon capture and greenhouse gas mitigation demand innovative strategies that harness biomolecular functions and integrate them into existing technologies. Enzyme-based systems offer a promising solution for sustainable CO₂ capture, yet their industrial adoption is limited by limited stability under harsh industrial conditions and the complexity of experimental optimization. Leveraging high-performance computing and large-scale molecular dynamics simulations, this study explores the design and behavior of biomimetic materials under diverse conditions, providing molecular insights to overcome these limitations. Carbonic anhydrase (CA), a ubiquitous metalloenzyme that efficiently converts CO₂ to bicarbonate—a fundamental physiological process in most living organisms—was selected as a model system for its remarkable catalytic performance and well-studied mechanism. Our sequence- and structure-based analyses uncovered critical factors such as pH, salt concentration, orientations and interactions of immobilized enzyme on surfaces. These insights, including the enzyme’s performance at variable pH, enabled the identification of strategies to improve catalytic efficiency and durability. By integrating these computational insights with experimental validation, this work establishes a foundation for robust, scalable CA-based systems for sustainable CO₂ capture, advancing global efforts to mitigate greenhouse gas emissions.
Author(s): Merve Fedai (North Carolina State University), and Yaroslava Yingling (North Carolina State University)
Development of a Predictive Model for the Prognosis of Patients with Breast Cancer
Breast cancer is one of the most prevalent malignancies among women, accounting for about a quarter of all new diagnoses worldwide. Treatment and prognosis vary according to histological subtype and stage at diagnosis. Because of heterogeneity in treatment responses, biomarkers that predict clinical outcomes are crucial. Recently, the neutrophil-to-lymphocyte ratio (NLR) has emerged as a promising prognostic biomarker. This study analyzed retrospective data from patients diagnosed with breast cancer between 2008 and 2022, focusing on complete blood count (CBC) findings and their prognostic impact. Absolute values and ratios (neutrophils, monocytes, basophils, platelets, and lymphocytes) were derived from CBCs performed before and throughout treatment. The results showed that NLR, monocyte/lymphocyte ratio (MLR), and platelet/lymphocyte ratio (PLR) serve as independent prognostic factors, while the basophil-to-lymphocyte ratio did not reach significance. These hematologic features, along with clinical data, were used to train supervised machine learning models classifying poor prognosis as recurrence or death within 10 years, and good prognosis otherwise.
Author(s): Patricia Honorato Moreira (Inteli – Institute of Technology and Leadership)
Distributed Computing for Spatio-Temporal Bayesian Modeling Using the INLA Method
Bayesian inference on large-scale spatio-temporal models is limited by its computational feasibility, a trend that is further exacerbated by the continuous increase in data availability and model refinements. To address this issue, we present a double-layer distributed-memory parallelization strategy for the popular integrated nested Laplace approximations (INLA) method. First, we perform in parallel the different objective function evaluations that happen during the hyperparameter minimization process. Each of these function evaluations requires the assembly and decomposition of large, often structured, sparse precision matrices. A second layer of parallelism is thus found in the decomposition of these sparse matrices. They are handled by a GPU-accelerated, distributed memory, direct solver called Serinv, which we have integrated into our framework. The latter, named pyINLA, is a novel Python package that aims at providing modern coding practices and features for spatio-temporal modeling within the INLA method. PyINLA outperforms the scalability of state-of-the-art packages, allowing the entire framework to run on distributed memory, GPU-accelerated, systems. We showcase the computational performance of our framework on the CSCS Alps supercomputer, scaling further a large spatio-temporal air temperature prediction model made of a spatial mesh composed of 2865 edges over the course of 365 days.
Author(s): Vincent Maillou (ETH Zurich), Alexandros Nikolaos Ziogas (ETH Zurich), Mathieu Luisier (ETH Zurich), Håvard Rue (King Abdullah University of Science and Technology), and Lisa Gaedke-merzhaeuser (King Abdullah University of Science and Technology, Università della Svizzera italiana)
Multi-Team Software Collaboration within the Exascale Computing Project
Collaboration and team science are emerging areas of interest in research and software production. Historically, multi-institutional research collaborations are difficult to initiate and maintain, negatively impacting communication, negotiation, and dialogue between industry, government, and academic researchers. The Exascale Computing Project was created to facilitate broader research collaboration under a shared funding structure and extended timeline to support scientific discovery. We conducted interviews with ECP teams, representing a variety of domain specialities, research institutions, programming backgrounds, and application areas. Using grounded theory as our analytical framework, we assessed how ECP’s structure created an environment of increased trust between projects and how software shared between teams facilitated sustained collaboration. Based on our findings within ECP projects, we share recommendations for sustainable multi-institutional collaboration and best shared software practices.
Author(s): Hana Frluckaj (University of Texas at Austin)
A Performance Portable Matrix-Free Finite Element Framework for Particle-Mesh Methods
Computing architectures are becoming increasingly complex and potent, as we reach new computing capacities. Currently the first three machines in the TOP500 list are exascale systems. To be able to take full advantage of these machines, and even run on such heterogeneous architectures, it has become imperative for computational simulations to be massively parallelized and hardware portable. The IPPL framework is a C++ library which allows users to write Particle-Mesh simulations in a performant and portable way, as it contains all the building blocks required for such simulations, such as meshes, particles, interpolation, and field solvers. We present a new solver implemented inside IPPL, a matrix-free FEM-based Poisson solver working from 1 to 3 dimensions, implemented in a fully portable way using Kokkos. The Finite Element framework implemented in IPPL is itself done in a modular way, allowing one to build on top of it to solve other PDEs. We showcase portability and performance by running scaling studies on the JEDI machine at Jülich Supercomputing Center, which has the new Nvidia GH200 GPUs, as well as the JURECA machine of the same center, with Nvidia A100 GPUs.
Author(s): Sonali Mayani (Paul Scherrer Institute, ETH Zurich)
Understanding HMM Performance for Enhanced HPC Portability
Heterogeneous Memory Management (HMM) simplifies programming for heterogeneous systems, making High-Performance Computing (HPC) devices more accessible to domain scientists; however, it suffers from slow performance compared to other memory management approaches. HMM is an infrastructure provided by Linux to enable a more simple and universal usage of non-conventional memory, enabling usage of multiple devices without the need for developers to use runtime APIs for memory allocation and data transfer. This simplification benefits domain scientists by reducing code complexity, making it easier to transition between systems, and enabling quicker adoption of legacy code for complex heterogeneous systems, such as those found in HPC Centers. Currently, HMM has slow performance compared to explicit memory management and Unified Virtual Memory (UVM), a similar infrastructure provided by NVIDIA for their GPUs. UVM requires driver specific APIs for allocation, but once allocated the memory can be used by any device. Due to the similarity between HMM and UVM, we expect any performance differences to result from improper UVM driver implementation, challenges in utilizing HMM correctly, or inefficient algorithms introduced by the abstraction. In our work, we conduct experiments to identify the root cause of the expected underlying issues and provide insights into their impact.
Author(s): Nicholas Cassarino (University of North Carolina at Charlotte)