The U.S. National Science Foundation today announced a new partnership with Nvidia Corp. to develop artificial intelligence models designed to advance scientific research across the country.

The collaboration will support the NSF Mid-Scale Research project known as Open Multimodal Infrastructure to Accelerate Science, or OMAI. The project will be led by the Allen Institute for AI, or Ai2, which will build a national, fully open AI ecosystem to accelerate both scientific discovery and the science of AI itself.

Ai2, well-known for its work on multimodal AI large language models, will bring its expertise to develop domain-specific LLMs trained on scientific literature. These models are intended to enable researchers to process and analyze research faster, generate code and visualizations and connect emerging insights with past discoveries.

As part of this public-private investment, NSF is contributing $75 million, and Nvidia is investing $77 million, all directed to Ai2 to lead this effort. The initiative aligns with priorities outlined in the White House AI Action Plan to deliver on AI-enabled science innovation and ensure the U.S. remains a global leader in open AI model development.

“This is a lasting national asset,” said Noah A. Smith, senior director of natural language processing research at Ai2 and professor at the University of Washington. “The AI infrastructure and the know-how that we’re building will advance both science and the science of AI, securing America’s leadership in innovation for decades to come,”

The OMAI models will be fully open, meaning their training data, code and evaluation methods will be transparent and reproducible — setting a new standard for openness in AI for science. Ai2 says it plans to continuously stream in new scientific literature to keep the models current, ensuring that they reflect the latest discoveries across fields.

Customization will also be central. Though general-purpose models work well for conversational use, scientific applications often require more domain-specific alignment.

“We at Ai2 are not believers in general-purpose, one-size-fits-all when it comes to AI,” said Smith. “We believe that customization is going to play a key role, especially in high-stakes areas like science.”

Nvidia will provide its cutting-edge HGX B300 systems, high-performance AI infrastructure designed to accelerate model training and deployment at scale, along with Nvidia AI Enterprise, the company’s cloud-native software tools, libraries and frameworks for AI development and deployment. This support is aimed at transforming massive datasets into actionable scientific intelligence faster than ever.

The NSF said the investment represents a first-of-its-kind leap in AI infrastructure for the foundation’s scientific mission.

“This marks the first time that NSF is making a major investment in the software infrastructure needed to power AI-driven science,” said Tess DeBlanc-Knowles, special assistant for artificial intelligence at NSF. “We see this investment as arriving at a critical moment. Because the question isn’t if AI will transform science, it’s who will lead that transformation.”

Ai2 will work with a scientific advisory committee to determine which research communities to prioritize in the early phases. Initial focus areas include accelerating the discovery of new materials, improving protein function prediction, and addressing known weaknesses in existing LLMs.

Though the full project timeline includes a major release around 18 months in, Ai2 plans to share key artifacts like datasets and evaluations throughout the development process to maintain transparency and community engagement.

Image: SiliconANGLE/Microsoft Designer