South Korea's Ministry of Science and ICT is moving to develop next-generation AI foundational technologies that predict scientific phenomena by using physical and mathematical principles, beyond data learning.

The ministry said on Monday it held a kickoff briefing for the "Next-Generation AI+Science and Technology (S&T) Foundational Technology Development Project" and will begin research and development on 4 newly selected tasks.

AI is currently used mainly by learning vast amounts of data to predict results. But in science and engineering fields where physical laws or mathematical principles are important, AI has limits because it is hard to explain why it produced a result and accuracy falls under new conditions.

The ministry will develop foundational technologies that enable AI to interpret and predict scientific phenomena more accurately and reliably by using physical and mathematical principles as well as data. It also plans to develop new AI architectures and training methods that go beyond the structural limits of existing AI models.

The project is split into 2 areas, including "physics and math-based AI interpretation and prediction models" and "next-generation AI architecture research". A total of 4 tasks were selected, with 2 in each area. The ministry will invest 20 billion won from 2026 to 2031. This year's budget is 2 billion won. It will support the selected researchers with computing infrastructure such as graphics processing units (GPUs). Research data secured and AI models developed during the project will be opened through a public platform.

In the area of physics and math-based AI interpretation and prediction models, the ministry will pursue development of AI models that operate stably even when conditions and environments change. A team led by Woo-seok Ha (하우석), a professor at the Korea Advanced Institute of Science and Technology (KAIST), will develop a math and physics-based causal AI model that infers causal structures and governing equations embedded in data. The goal is to produce results based on physical laws even when conditions and environments change.

A team led by Jae-seok Yoo (유재석), a professor at Daegu Gyeongbuk Institute of Science and Technology (DGIST), will develop an "AI Mechanician" that visualises which physical laws act dominantly across various phenomena and enables AI to choose interpretation strategies on its own.

The ministry expects the technologies to reduce the process of validating models from scratch each time conditions change, cutting the cost and time needed for design, interpretation and validation.

In the area of next-generation AI architecture research, work will be carried out to identify the learning and structural principles of existing AI models and design new architectures. A team led by Min-hwan Oh (오민환), a professor at Seoul National University, will mathematically identify the structural limitation in which computation rises sharply as context gets longer, and will develop new AI architectures and training methods that can handle long contexts more efficiently.

A team led by Cheol-hee Yoon (윤철희), a KAIST professor, will mathematically identify the principles by which generative AI models improve performance during training. Based on that, it plans to establish next-generation AI development methodologies to apply across the entire process, from model architecture design to training and reliability verification.

Gyeong-suk Yoon (윤경숙), director general for Basic and Fundamental Research Policy at the ministry, said, "Securing new AI model architectures could also contribute to the advancement of AI technology itself."