Elements: ForteFAI - Trustworthy AI Cyberinfrastructure for Guardrailing Multimodal Foundation Models and Cyberinfrastructure

ForteFAI strengthens the reliability of artificial intelligence systems that increasingly serve as critical cyberinfrastructure for science and engineering. Foundation models are now used for scientific imaging, multimodal data analysis, natural-language interfaces, and operational monitoring, yet they often fail silently when confronted with data that differ from those encountered during training. ForteFAI addresses this challenge by providing a unified framework for detecting distributional anomalies across vision, language, and cyberinfrastructure domains. The project develops open, production-ready software that improves quality control for scientific imaging, enhances safeguards for large language model workflows, and enables anomaly detection in high-performance computing operations. By identifying unreliable inputs before they propagate through downstream analyses, ForteFAI improves the robustness, reproducibility, and trustworthiness of artificial intelligence–enabled scientific discovery. The project advances trustworthy artificial intelligence through FORTE (Finding Outliers with Representation Typicality Estimation), a model-agnostic methodology that quantifies how closely an input aligns with the representation manifold of trusted data. FORTE estimates semantic typicality using geometric manifold statistics derived from foundation model embeddings, including precision, recall, density, and coverage metrics, enabling unsupervised detection of distributional shifts without labeled anomalies, likelihood modeling, or domain-specific retraining. The research integrates this approach into three coordinated components: ForteFAI Vision for scientific image quality assurance and out-of-distribution detection; ForteFAI Text for large language model guardrails, hallucination detection, and prompt-injection monitoring; and ForteFAI Log for parser-free anomaly detection in high-performance computing logs. The resulting software, training materials, and community engagement activities will broaden access to trustworthy artificial intelligence capabilities and strengthen the national cyberinfrastructure ecosystem. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2608771 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Vipin Chaudhary | Institution: Case Western Reserve University, CLEVELAND, OH | Award Amount: $600,000 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2608771 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2608771.html