{"id":84736,"date":"2025-10-17T17:23:05","date_gmt":"2025-10-17T17:23:05","guid":{"rendered":"https:\/\/www.newsbeep.com\/nz\/84736\/"},"modified":"2025-10-17T17:23:05","modified_gmt":"2025-10-17T17:23:05","slug":"osc-research-earns-top-honors-at-annual-pearc-conference","status":"publish","type":"post","link":"https:\/\/www.newsbeep.com\/nz\/84736\/","title":{"rendered":"OSC Research Earns Top Honors at Annual PEARC Conference"},"content":{"rendered":"

COLUMBUS, Ohio, Oct. 17, 2025 \u2014\u00a0At the\u00a0Practice and Experience in Advanced Research Computing (PEARC) conference<\/a>, researchers gather each summer to share innovations in high performance computing (HPC), data, and advanced applications. This year, when\u00a0PEARC25<\/a>\u00a0came to Columbus, Ohio, two Ohio Supercomputer Center<\/a> (OSC) powered projects stood out\u2014earning top national awards<\/a>.<\/p>\n

Optimizing Machine Learning in the Field<\/p>\n

OSC Senior Research Software Engineer Samuel Khuvis was part of a multi-institutional team that earned both the Best Full Paper Award (Applications & Software Track) and the Phil Andrews Award for \u201cML Field Planner: Analyzing and Optimizing ML Pipelines for Field Research<\/a>.\u201d<\/p>\n

\"\"<\/a>Ohio Supercomputer Center researchers Samuel Khuvis (at left, left photo) and Ron Davies (at left, right photo) were recognized at the 2025 Practice and Experience in Advanced Research Computing (PEARC) conference for award-winning projects in machine learning and AI security. Credit: OSC.<\/p>\n

The ML Field Planner tool was designed for scientists who need to run machine learning outside of the data center. Ecologists, for example, often place camera traps (motion- and heat-sensitive self-activating cameras) that capture images or videos of wildlife without human interference. These traps rely on small edge devices like Raspberry Pis or Jetson Nanos, which have limited battery life, storage, and processing power. Running the wrong machine learning model on these devices can waste energy and storage\u2014or miss important images entirely.<\/p>\n

ML Field Planner helps researchers test those choices before heading into the field. Through a web-based dashboard, they can run experiments on real hardware to compare models side by side, measure accuracy against specific datasets, and monitor how much energy each model consumes.<\/p>\n

At\u00a0The Wilds Conservation Center<\/a>\u00a0in Ohio, ecologists used the ML Field Planner tool to evaluate versions of MegaDetector, a widely used machine learning model that scans camera-trap images and flags whether they contain animals, humans, or vehicles.<\/p>\n

The comparison focused on two versions:<\/p>\n

MegaDetector v5 \u2014 larger and more powerful, but slower and more energy-intensive.
\nMegaDetector v6 \u2014 smaller, faster, and more efficient, designed for edge devices.<\/p>\n

Results showed that v5 initially performed better on small animal images, but when fine-tuned with local data it became too narrow, losing accuracy. Meanwhile, v6 improved with fine-tuning and proved to be the better option for long-term use in the field.<\/p>\n

By making those comparisons clear, ML Field Planner gave ecologists confidence in how to deploy their devices: Start with v5 for short-term reliability, but transition to v6 as more field data becomes available.<\/p>\n

\u201cIt really lets people see tradeoffs in advance\u2014which model is faster, which one uses less energy, which one is worth deploying,\u201d Khuvis said.<\/p>\n

And while it was tested on wildlife monitoring, the same approach could help scientists in agriculture, climate studies, and other environmental research who face the same challenge: bringing machine learning out of the lab and into the field.<\/p>\n

The project is part of the National Science Foundation-funded\u00a0ICICLE AI Institute<\/a>, led by The Ohio State University with OSC as a key partner. ML Field Planner was developed within the ICICLE software stack and integrated into the ICICLE ML WorkBench, which provides researchers with advanced tools for applying artificial intelligence to complex, real-world problems.<\/p>\n

Making AI Interfaces Safer <\/p>\n

Ohio State Computer Science and Engineering<\/a>\u00a0doctoral student Ron Davies was part of a team recognized with the Best Student Poster Award for \u201cA Pre-Processing Framework for Securing LLM-RAG Interfaces Against Information Leakage<\/a>.\u201d The project grew out of Davies\u2019 work with the\u00a0Open OnDemand<\/a>\u00a0community, OSC\u2019s flagship web portal that gives thousands of researchers an easy way to connect to supercomputing resources through a simple online interface.<\/p>\n

As artificial intelligence (AI) becomes more integrated into HPC, Davies and his co-authors saw both an opportunity and a risk. Large language model (LLM) tools like ChatGPT or Microsoft Copilot make it easier for users to interact with powerful systems, but they also can open the door for security problems. A poorly phrased or malicious prompt could waste expensive computing time or, worse, leak sensitive information about the underlying infrastructure.<\/p>\n

The team\u2019s solution was a lightweight, controlling path that acts as a checkpoint before queries ever reach a large, resource-intensive language model. The framework uses:<\/p>\n

A smaller screening model that quickly evaluates whether a query is safe and relevant.
\nCrafted memories\u2014curated dialogue examples showing safe vs. unsafe queries\u2014to help the model recognize what should and shouldn\u2019t pass through.
\nReminder prompts that reinforce boundaries, such as \u201conly answer questions about supercomputing tasks.\u201d<\/p>\n

In testing, the framework proved effective, blocking more than 90% of irrelevant or potentially harmful queries while still letting legitimate research questions through. By filtering at the front door, the system protects both security and efficiency\u2014ensuring valuable computing power is spent only on real research.<\/p>\n

For Open OnDemand, the implications are significant. This framework could support a chatbot module that gives users a low-stakes, conversational way to access supercomputing resources without fear of breaking something or exposing sensitive data. That balance\u2014ease of use plus strong security\u2014could make advanced computing more approachable to a wider range of researchers.<\/p>\n

Davies credits his involvement with Open OnDemand for shaping his approach.<\/p>\n

\u201cIt\u2019s given me a powerful perspective I haven\u2019t had before,\u201d he said, noting that sitting in on Open OnDemand leadership meetings helped him understand how central teams and community developers work together. He also pointed to his experience mentoring high school students at\u00a0OSC\u2019s STEM Institute<\/a>\u00a0as a way to see firsthand where new users struggle.<\/p>\n

Together, the two projects involving Khuvis and Davies highlight the breadth of OSC\u2019s impact\u2014from empowering students to design secure AI tools to enabling research teams to optimize machine learning for real-world fieldwork. Recognition at PEARC25 affirms that with accessible resources, strong collaboration, and advanced infrastructure, OSC is helping researchers push boundaries today while shaping the future of research computing.<\/p>\n

About OSC<\/p>\n

The Ohio Supercomputer Center<\/a> (OSC) addresses the rising computational demands of academic and industrial research communities by providing a robust shared infrastructure and proven expertise in advanced modeling, simulation and analysis. OSC empowers scientists with the services essential to making extraordinary discoveries and innovations, partners with businesses and industry to leverage computational science as a competitive force in the global knowledge economy and leads efforts to equip the workforce with the key technology skills required for 21st century jobs.<\/p>\n

Source<\/a>: Lexi Biasi, OSC<\/p>\n","protected":false},"excerpt":{"rendered":"COLUMBUS, Ohio, Oct. 17, 2025 \u2014\u00a0At the\u00a0Practice and Experience in Advanced Research Computing (PEARC) conference, researchers gather each…\n","protected":false},"author":2,"featured_media":84737,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[21],"tags":[371,111,139,69,145],"class_list":{"0":"post-84736","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-computing","8":"tag-computing","9":"tag-new-zealand","10":"tag-newzealand","11":"tag-nz","12":"tag-technology"},"_links":{"self":[{"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/posts\/84736","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/comments?post=84736"}],"version-history":[{"count":0,"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/posts\/84736\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/media\/84737"}],"wp:attachment":[{"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/media?parent=84736"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/categories?post=84736"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.newsbeep.com\/nz\/wp-json\/wp\/v2\/tags?post=84736"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}