South Korea’s Ministry of Science and ICT unveiled the Next-Generation AI+S&T Foundational Technology Development Project, a six-year effort from 2026 to 2031 with a total budget of 20 billion won, including 2 billion won for the current year.

The project aims to develop physics- and math-based AI that predicts scientific phenomena using fundamental principles alongside data, pushing beyond traditional data-driven learning.

Leading scientists will drive the physics and math-based track: Woo-seok Ha of KAIST will lead a causal AI model, and Jae-seok Yoo of DGIST will spearhead an AI Mechanician that identifies governing physical laws across phenomena.

The program comprises two research areas with four total tasks: physics and math-based interpretation and prediction, and next-generation AI architecture research.

Top policy leadership, including Gyeong-suk Yoon, stressed that securing new AI architectures is essential for advancing AI technology itself.

Researchers will gain access to a public platform for research data and AI models, supported by computing infrastructure such as GPUs.

Two architecture teams will drive the AI architecture track: Min-hwan Oh from Seoul National University will address reducing computational blow-up with longer contexts, and Cheol-hee Yoon from KAIST will pursue principled training improvements with cross-cutting applicability.

The architecture track will explore longer-context handling and training dynamics designed to boost generative AI performance, with aims to develop new architectures and training methods.

In the physics-and-math track, teams will focus on stability across changing conditions, causal AI to infer governing equations from data, and visualization of dominant physical laws to guide interpretation.

The initiative will publish research data and models on a public platform and will back researchers with GPU-enabled computing resources.