[대학원 세미나] 10/16(금) 김남중(서강대학교 기계공학과) "A Prompt-based Foundation Model for Mechanical Metamaterial Design"

기계공학부 구성원들의 많은 관심과 참여 부탁드립니다.

▣ 제 목: A Prompt-based Foundation Model for Mechanical Metamaterial Design

▣ 연 사: 김남중 교수

▣ 소 속: 서강대학교 기계공학과

▣ 일 시: 2026. 10. 16.(금) 13:00

▣ 장 소: 제2공학관 B040호

▣ 초 록

Mechanical metamaterials—engineered materials whose properties emerge from their geometric architecture rather than their composition—exhibit extraordinary behaviors, including programmable stiffness, negative Poisson’s ratios, and extreme energy absorption. However, discovering new architectures remains challenging due to the enormous design space and the cost of high-fidelity simulations. Recent advances in artificial intelligence (AI) have accelerated metamaterial design through surrogate modeling and generative approaches, yet most frameworks remain task-specific and require retraining for new problems. These limitations motivate generalizable learning paradigms capable of transferring knowledge across diverse metamaterial systems.

In this talk, I will present our recent advances in developing foundation models for mechanical metamaterials, inspired by the success of language and vision foundation models. Central to this effort is the Metamaterial Foundation Operator (MetaFO), a prompt-based framework that learns generalized mappings between material properties, structural architectures, and mechanical responses. Through in-context learning, MetaFO enables zero-shot prediction and probabilistic reasoning for unseen material–structure combinations without task-specific retraining, while maintaining robustness under nonlinear and out-of-distribution conditions.

Beyond the core foundation model, I will discuss emerging directions toward improved scalability, efficiency, and generalization. These include compressed and quantized foundation models for resource-efficient deployment, mixture-of-experts architectures for capturing diverse mechanical behaviors, and extensions to multimaterial metamaterial systems. Together, these advances establish a new paradigm for data-efficientlearning and knowledge transfer in mechanics-driven design.