The rapid growth of artificial intelligence, simulation, and data-intensive science has created new opportunities across nearly every research field. At the same time, many promising projects require access to specialized GPU computing resources and expertise that may be difficult for individual research groups to obtain. MIT's Office of Research Computing and Data (ORCD) created its Seed Fund program is part of MIT's strategic effort to advance AI and machine learning research by providing researchers with access to GPU-accelerated computing resources and technical collaboration from ORCD staff. The program supports short-term, high-impact projects that both accelerate discovery and help evaluate and improve MIT's shared research computing environment.

The Fall 2026 ORCD Seed Fund portfolio spans materials science, astronomy, applied mathematics, fusion energy, biodiversity monitoring, biomedical imaging, neurology, and artificial intelligence.

• Computational Prediction of Phase Diagrams for Metallic Alloys in Silico uses machine learning and GPU-accelerated simulations to predict alloy behavior across temperatures and compositions, accelerating materials discovery.
• Developing a Public Package for Image Segmentation of Astronomical Datasets applies modern AI segmentation methods to protoplanetary disk imagery, helping astronomers more accurately interpret telescope observations.
• Multi-scale Mixed-Precision Hardware-Agnostic Linear Algebra seeks to make advanced numerical computations portable across computing architectures while maintaining performance.
• Machine-Learning Enabled Predictions of Fusion Power Plant Performance combines high-performance simulations and machine learning to improve predictions of future fusion reactor behavior.
• Generative AI with Environmental Context Matching for Rare Insect Classification develops synthetic training data that reflects real-world ecological conditions, improving species-identification tools.
• MultiverSeg3D aims to enable scalable segmentation of large biomedical imaging datasets using foundation-model approaches.
• Multimodal Brain Foundation Model for Neurology combines EEG and fMRI data to advance neurological research and seizure forecasting.
• Look but Don't Touch: Gradient Informed Selection Training develops more efficient AI training methods that can reduce computational cost while improving model performance.

Across all projects, MIT research computing provides the GPU infrastructure, software environment, and technical expertise that enable researchers to tackle computational problems at scales that would otherwise be impractical.