|
1 |
Design Foundation Model (DFM)
DFM team develops AI-driven design technologies for engineering practice, which require both generalization across a broad range of engineering design problems and automated verification of generated designs. To this end, our research aims to develop a design foundation model applicable to diverse design problems, together with a verification agent that automatically validates generated designs against engineering-design references and engineer’s accumulated experience.
Foundation Model for Engineering Design
We develop foundation models that transfer knowledge from broad design and simulation datasets to new engineering problems, enabling inverse design and rapid performance assessment with less task-specific data and training.
Inverse Design with Design Foundation Models
- MOTIVATION
- Conventional data-driven design models are trained for specific products and parameterizations, forcing every new task to start from scratch without reusing accumulated design knowledge.
- APPROACH
- Design foundation models are developed to generate candidates in a single pass from the required performance, where a tabular model infers a distribution over design variables in context and a topology model generates the geometry itself.
- IMPACT
- Pretrained once, the models transfer to a new system with a few examples instead of a new full training process, and the tabular and topology branches extend to a single multimodal design foundation model that handles parameters and geometry together.
Design Verification with Analysis Foundation Model
- MOTIVATION
- AI-generated designs are not guaranteed to satisfy physical requirements, yet conventional surrogates are bound to one geometry and one physics setting, so verification cannot keep up with generation.
- APPROACH
- An analysis foundation model is pretrained on diverse simulation data and fine-tuned for target structures, enabling real-time prediction of physics fields, performance, and uncertainty for candidate designs.
- IMPACT
- A single pretrained model verifies designs across geometries and physics without building a task-specific surrogate, screening generated candidates so that only reliable ones proceed to full simulation.
Knowledge-grounded Agent for Design Verification
We develop an ontology-based agent that integrates engineering references and tacit expert knowledge to verify designs against requirements and constraints, providing traceable, evidence-based feedback for design improvement.
Design-Verification Agent guided by Ontology
- MOTIVATION
- AI-generated designs are not guaranteed to satisfy engineering requirements, and their verification depends on engineers’ knowledge scattered across heterogeneous sources.
- APPROACH
- A design-verification agent is developed guided by heterogeneous ontology that integrates engineering-design references and engineers’ experience.
- IMPACT
- The agent enables automated, reliable verification of AI-generated designs, accelerating the adoption of AI in engineering design.
|
