BEGIN:VCALENDAR VERSION:2.0 PRODID:Linklings LLC BEGIN:VTIMEZONE TZID:America/Chicago X-LIC-LOCATION:America/Chicago BEGIN:DAYLIGHT TZOFFSETFROM:-0600 TZOFFSETTO:-0500 TZNAME:CDT DTSTART:19700308T020000 RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU END:DAYLIGHT BEGIN:STANDARD TZOFFSETFROM:-0500 TZOFFSETTO:-0600 TZNAME:CST DTSTART:19701101T020000 RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20260202T201319Z LOCATION:241 DTSTART;TZID=America/Chicago:20251116T090000 DTEND;TZID=America/Chicago:20251116T173000 UID:submissions.supercomputing.org_SC25_sess209@linklings.com SUMMARY:Frontiers in Generative AI for HPC Science and Engineering: Founda tions, Challenges, and Opportunities DESCRIPTION:Realizing the promise of large-scale foundation models for sci entific discovery—enabling self-driving laboratories, hypothesis generatio n, and more—requires unprecedented computational scale and multidisciplina ry efforts to prepare diverse scientific data. While only a few organizati ons can train state-of-the-art models from scratch (e.g., trillions of par ameters, tens of trillions of tokens), advances in training strategies and fine-tuning have expanded accessibility. Simultaneously, breakthroughs in training methodologies and data quality are dramatically reducing trainin g costs and improving the performance of even smaller AI models. As AI mod els advance in general-purpose tasks, the scientific community is refining methods to evaluate and enhance their scientific reasoning capabilities, a critical challenge for trustworthy AI in science. This workshop, catalyz ed by the Trillion Parameter Consortium (TPC), will highlight collaboratio ns in scientific skills evaluation, performance optimization, federated le arning, responsible AI, and other topics. SC24 drew 33 submissions, with 1 3 presented to nearly 200 attendees, underscoring the rapid evolution of t his field.\n\nAfternoon Break - Frontiers in Generative AI for HPC Science and Engineering: Foundations, Challenges, and Opportunities\n------------ ---------\nAn Update on TPC-Coordinated Global Collaborative Projects\n\nT PC Leaders will discuss progress on organizing and incubating collaborativ e TPC initiatives that will drive hackathons in 2026, including how these initiatives inter-relate, a strategy for ensuring that they are applicatio n-driven, and how to get involved.\n\n---------------------\nAutomated MCQ A Benchmarking at Scale: Evaluating Reasoning Traces as Retrieval Sources for Domain Adaptation of Small Language Models\n\nAs scientific knowledge grows at an unprecedented pace, evaluation benchmarks must evolve to refle ct new discoveries and ensure language models are tested on current, diver se literature. We propose a scalable, modular framework for generating mul tiple-choice question-answering (MCQA) benchmarks dire...\n\n\nOzan Gokdem ir (University of Chicago, Argonne National Laboratory (ANL)) and Neil Get ty, Robert Underwood, Sandeep Madireddy, Franck Cappello, Arvind Ramanatha n, Ian Foster, and Rick Stevens (Argonne National Laboratory (ANL))\n----- ----------------\nBatch Tiling on Attention: Efficient Mixture of Experts Training on Wafer-Scale Processors\n\nMixture of Experts (MoE) models face computational bottlenecks on wafer-scale processors due to conflicting ba tch size requirements: attention mechanisms need smaller batches for memor y constraints, while routable MLP layers require larger batches for optima l compute density.\nWe introduce \textbf{Ba...\n\n\nDaria Soboleva, Etienn e Goffinet, Hui Zeng, Sangamesh Ragate, Elif Albuz, and Natalia Vassilieva (Cerebras)\n---------------------\nBioR5: A Three-Layer Architecture for Biological Reasoning in Scientific AI\n\nBiological research requires dive rse reasoning modes, from phylogenetic analysis to mechanistic understandi ng, each demanding specific methods and data types. Current AI systems typ ically employ single methodologies, limiting effectiveness in complex biol ogical domains. We present BioR5, a three-lay...\n\n\nPeng Ding (Universit y of Chicago, Argonne National Laboratory (ANL)); Thomas Brettin (Argonne National Laboratory (ANL)); and Rick Stevens (Argonne National Laboratory (ANL), University of Chicago)\n---------------------\nChatEED: An agentic retrieval assistant for accelerator operators\n\nScientists and operators at SLAC National Accelerator Laboratory rely on electronic logbooks (ELOGs ) to record and share information surrounding accelerator operations. Howe ver, since creating log entries is time-consuming and complex, they are of ten brief, incomplete, jargonized, and inconsistent. ...\n\n\nAaron Reed, Claudio Bisegni, Sandesh Shrestha, Michelle Huang, and Daniel Ratner (SLAC National Accelerator Laboratory)\n---------------------\nGridMind: LLMs-P owered Agents for Power System Analysis and Operations\n\nThe complexity o f traditional power system analysis workflows presents significant barrier s to efficient decision-making in modern electric grids. This paper presen ts GridMind, a multi-agent AI system that integrates Large Language Models (LLMs) with deterministic engineering solvers to enable conve...\n\n\nHon gwei Jin, Kibaek Kim, and Jonghwan Kwon (Argonne National Laboratory (ANL) )\n---------------------\nEQSIM Agent: A Conversational AI for Interactive Exploration of Large-scale Earthquake Simulation Data\n\nLarge-scale eart hquake simulations produce massive, high-fidelity datasets essential for s eismic risk analysis, however, their volume and complexity create a barrie r for researchers from various backgrounds who lack specialized knowledge and programming skills. To address this challenge, we leverage...\n\n\nHou jun Tang and David McCallen (Lawrence Berkeley National Laboratory (LBNL)) \n---------------------\nEvaluation of Test-Time Compute Constraints on Sa fety and Skill Large Reasoning Models\n\nTest-time compute scaling has dem onstrated the ability to improve the performance of reasoning language mod els by generating longer chain-of-thought (CoT) sequences. However, this i ncrease in performance comes with a significant increase in computation co st. In this work, we investigate two compute ...\n\n\nAdarsha Balaji, Le C hen, Rajeev Thakur, Franck Cappello, and Sandeep Madireddy (Argonne Nation al Laboratory (ANL))\n---------------------\nFrameworks for Large Language Model Serving in HPC Environments\n\nWe introduce open-source frameworks for deploying and running large language models (LLMs) within high-perform ance computing (HPC) environments. One such framework targets high-\nthrou ghput batch inference, enabling users to submit LLM requests in an OpenAI- compatible format as traditional HPC jobs. ...\n\n\nRohan Marwaha (Univers ity of Illinois at Urbana-Champaign); Qinren Zhou (University of Illinois at Urbana-Champaign, National Center for Supercomputing Applications (NCSA )); Kastan Day and Asmita Dabholkar (University of Illinois at Urbana-Cham paign); and Volodymyr Kindratenko (University of Illinois at Urbana-Champa ign, National Center for Supercomputing Applications (NCSA))\n------------ ---------\nFrontiers in Generative AI for HPC Science and Engineering: Fou ndations, Challenges, and Opportunities\n\nRealizing the promise of large- scale foundation models for scientific discovery—enabling self-driving lab oratories, hypothesis generation, and more—requires unprecedented computat ional scale and multidisciplinary efforts to prepare diverse scientific da ta. While only a few organizations ...\n\n\nCharlie Catlett (Argonne Natio nal Laboratory (ANL), University of Chicago); Fabrizio Gagliardi (Barcelon a Supercomputing Center (BSC)); Valerie Taylor (Argonne National Laborator y (ANL), Univeristy of Chicago); Laura Morselli (CINECA); and Rio Yokota ( Institute of Science Tokyo)\n---------------------\nBeyond End-to-End: Und erstanding the Limits of LLMs in Scientific Problem Solving\n\nMultimodal large language models (MLLMs) are now widely used across many applications , including scientific question answering that requires combining visual a nd textual inputs. However, existing benchmarks in this area are mostly en d-to-end, making it difficult to pinpoint where models fail. To add...\n\n \nYouyuan Liu (Temple University); Sheng Di, Neil Getty, Tanwi Mallick, an d Robert Underwood (Argonne National Laboratory (ANL)); and Sian Jin (Temp le University)\n---------------------\nExploring Distributed Vector Databa ses Performance on HPC Platforms: A Study with Qdrant\n\nVector databases have rapidly grown in popularity, enabling efficient similarity search ove r data such as text, images, and video. They now play a central role in mo dern AI workflows, aiding large language models by grounding model outputs in external literature through retrieval-augmented generatio...\n\n\nSeth Ockerman (University of Wisconsin-Madison, Argonne National Laboratory (A NL)); Amal Gueroudji (Argonne National Laboratory (ANL)); Song Young Oh (U niversity of Chicago); Robert Underwood and Nicholas Chia (Argonne Nationa l Laboratory (ANL)); Kyle Chard (University of Chicago, Argonne National L aboratory (ANL)); Robert Ross (Argonne National Laboratory (ANL)); and Shi varam Venkataraman (University of Wisconsin-Madison)\n-------------------- -\nLABMATE: Language Model Based Multi-Agent System to Accelerate Catalysi s Experiments\n\nLarge Language Models (LLMs) are unreliable to make decis ions due to their potential to hallucinate, and unable to perform complex tasks like running simulations essential to fields like Material Science. We introduce LABMATE (LAnguage model Based Multi-agent system to Accelerat e caTalysis Experimen...\n\n\nAnurag Acharya (Pacific Northwest National L aboratory (PNNL)); Anshu Kiran Sharma (Florida International University, P acific Northwest National Laboratory (PNNL)); and Derek Parker, Timothy Ve ga, Rizwan A. Ashraf, Natalie M. Isenberg, Jan Strube, and Robert Rallo (P acific Northwest National Laboratory (PNNL))\n---------------------\nAgent ic AI vs ML-based Autotuning: A Comparative Study for Loop Reordering Opti mization\n\nHigh Performance Computing (HPC) applications rely heavily on code optimizations to achieve good performance on modern CPU and GPU arch itectures. Traditional Machine Learning autotuning approaches have demonst rated success in exploring high-dimensional spaces, but they often require expensive compi...\n\n\nMiguel Romero Rosas (University of Delaware), Kha led Ibrahim (Lawrence Berkeley National Laboratory (LBNL)), and Rudolf Eig enmann (University of Delaware)\n---------------------\nWorkshop Welcome a nd Overview\n\nWorkshop organizers will provide an overview of the objecti ves and program as well as the international Trillion Parameter Consortium and opportunities to engage.\n\n\nJavier Aula-Blasco (Barcelona Supercomp uting Center (BSC))\n---------------------\nMorning Break - Frontiers in G enerative AI for HPC Science and Engineering: Foundations, Challenges, and Opportunities\n---------------------\nLunch break (on your own)\n\nRecord ing: Livestreamed, Recorded\n\nRegistration Category: Technical Program Re g Pass, Workshop Reg Pass\n\nSession Chairs: Charlie Catlett (Argonne Nati onal Laboratory (ANL), University of Chicago); Javier Aula-Blasco (Barcelo na Supercomputing Center (BSC)); Valerie Taylor (Argonne National Laborato ry (ANL), University of Chicago); Laura Morselli (CINECA); and Rio Yokota (Institute of Science Tokyo) END:VEVENT END:VCALENDAR