Microsoft Support for the National AI Research Resource Pilot
Grand Challenges
We’re excited by breakthroughs that can be enabled by leveraging state of the art models for health, molecular and materials discovery. Building on the momentum of breakthroughs to date using these models, Microsoft is eager to share the tools we have for advancing world-class science to support grand challenge research projects through the NAIRR Pilot’s new Deep Partnership track.
Grand Challenge Awards
We are pleased to recognize the following researchers who were selected as finalists for Azure grand challenge awards. See below for more information on their proposed projects.
Accelerating Multifunctional C-S-H Seeding Materials Discovery through Agentic AI and Scalable Cloud Computing
Prasanna Venkataraman Balachandran (University of Virginia)
Concrete is the most widely used man-made material globally, with its composition and production methods largely unchanged for over 200 years. Cement production is responsible for nearly 8% of total CO2 emissions. In developed nations, infrastructure repair is a major challenge, and the U.S. alone requires $2.8 trillion for its aging transportation systems. Low-carbon cements could reduce cement emissions by 20-60%, but they struggle with early-age strength development, hindering their adoption. Research on calcium-silicate-hydrate (C-S-H) seeding has shown promise in accelerating cement hydration, highlighting the need for new multifunctional C-S-H seeding materials. This proposal aims to develop an innovative AI pipeline that integrates advanced AI techniques with domain-specific constraints and physics-based simulations to discover and optimize novel C-S-H seeding materials.
Accelerating molecular design and crystal structure prediction for carbon capture
Jeffrey Neaton (University of California, Berkeley)
Crystal structure prediction (CSP) is a longstanding challenge in materials science, with conventional methods relying on computationally expensive density functional theory (DFT) calculations that limit their applicability to small systems. Recent advances in machine learning interatomic potentials (MLIPs) and generative AI offer a paradigm shift, enabling efficient exploration of vast configuration spaces with near-DFT accuracy at greatly reduced cost. This project aims to develop an integrated AI-driven CSP workflow that combines MLIPs, generative AI, DFT, and classical simulations to predict the crystal structures of complex organic systems, demonstrating our workflow on polyamine molecular crystals as promising candidates for carbon capture. The resulting workflow and transferable MLIPs will be made publicly available, providing powerful new tools for CSP of molecular crystals and accelerating the discovery of materials for energy and environmental applications.
RxFM: A Multimodal Foundation Model for ALS Therapeutic Science and Drug Discovery
Mark Albers (Massachusetts General Hospital, Harvard University)
Amyotrophic Lateral Sclerosis (ALS) is a fatal neurodegenerative disease with no cure and a complex, multifactorial etiology. This project aims to develop a Multimodal Therapeutic Foundation Model (RxFM) that will be trained on diverse biomedical and pharmacodynamic data to accelerate Amyotrophic Lateral Sclerosis (ALS) drug discovery and drug repurposing. This RxFM model will integrate multi-omic, single-cell, metabolomic, clinical, and drug perturbation data within a unified framework. The project aims to extend advances in biologic foundation models to model therapeutic actions, seeking to identify interventions that reverse disease signatures. Moreover, the RxFM framework can be fine-tuned or queried for diseases with little known biology, supporting zero-shot drug discovery.
Accelerating Neurodegeneration Discovery with Agentic AI Systems in C-BRAIN
Randall Bateman (Washington University in St. Louis)
Alzheimer’s disease research faces a critical challenge: despite decades of effort, more than 99% of drug candidates fail in clinical trials. This failure reflects a deeper problem in biomedical science – vital knowledge is scattered across millions of papers, complex datasets, and unpublished “dark data,” making it nearly impossible for a scientist to understand the complete picture.
The Consortium for Biomedical Research and AI in Neurodegeneration (C-BRAIN) is tackling this challenge by combining the expertise of leading scientists with cutting-edge AI. This project aims to develop and test three unique AI tools that mirror the way science works: generating hypotheses, testing them against data, and providing rigorous peer-review style critique. At each step, human experts will remain in the loop to guide, validate, and refine outputs, ensuring trust and scientific rigor. Together, these AI tools will form the first open benchmark for “AI Scientist” systems; thereby, allowing the broader research community to compare and improve their own approaches. The goal is to deliver a transparent, end-to-end pipeline that will accelerate discovery in Alzheimer’s and set the stage for scalable, AI-enabled science across biomedicine.
Braingents: Developing a Longitudinal Agentic AI Framework for Alzheimer’s Disease Neuroimaging Biomarkers
Ish Talati (Stanford University School of Medicine)
This project aims to develop an agentic AI framework that autonomously orchestrates end-to-end neuroimaging analysis while preserving expert-level accuracy. The framework will bring together four specialized agents: a Data Harmonization Agent that normalizes scanner and protocol variability across sites using domain adaptation; a Multi-Modal Detection Agent leveraging vision transformers to quantify structural atrophy and amyloid/tau burden; a Longitudinal Tracking Agent for precise monitoring of hippocampal volumes, cortical thinning, and plaque progression; and a Clinical Reasoning Agent that generates interpretable, language model–driven reports. By combining automation, interpretability, and multimodal scalability, the research aims to accelerate biomarker discovery while paving the way for clinical readiness in Alzheimer’s disease neuroimaging.
Trustworthy and Secure Cloud-Native Agentic OpenROAD
Andrew Kahng (University of California San Diego)
This project aims to improve the efficacy, trust, and security of agentic systems applied to integrated-circuit electronic design automation (IC EDA). The team will develop OpenROAD V2, a cloud-native, AI-assisted extension of the DARPA-funded OpenROAD IC EDA tool. OpenROAD V2 will expose design and code “knobs” in a controlled, declarative way, allowing LLM-guided agents to tune tool behavior without unsafe arbitrary edits. These innovations will enable runtime improvements of 3-10X and design outcome improvements of 15-20%, while providing a trusted “sandbox” for AI agents to self-adapt OpenROAD flows to specific technologies and designs. This work will establish a foundation for ML-driven and agentic EDA research, opening new opportunities in chip design research, education and industrial practice.
Towards Agentic LLM for Chiplet Design
Jun Zhang (Arizona State University)
This project aims to build an agentic AI model for end-to-end electronic design automation (EDA), spanning HDL generation, verification, and physical layout. The framework will build upon Reinforcement Learning from Internal Feedback (RLIF), which learns chip design tasks without the need of external labels. The approach is promising for chip design, a domain with scarce open-source data. By fine-tuning base models with RLIF and integrating verification agents and back-end tasks, the team plans to develop and automate workflows across the design stack with a unified, open-source EDA toolchain. The framework will further enable power-performance optimizations from natural language interfaces.