BROWSE
Detail
Journal Article
Idea Generation in AI for Science: A Survey from Hypothesis Genera-tion to Closed-Loop Discovery
- Authors
- Jin-Xia Huang, Woojin Lee, Yoonkyu Woo
- Issue Date
- 2026-06
- Citation
- InUS Transactions, v.2, no.2, pp.59-80
- ISSN
- 3092-0043
- Publisher
- 무인시스템학회
- Language
- English
- Type
- Journal Article
- Abstract
- Artificial intelligence increasingly proposes hypotheses, candidate structures, and experimental interventions within scientific discovery workflows. This survey examines \emph{idea generation} in AI for Science as the computational function that produces new candidates (textual hypotheses, symbolic equations, molecular graphs, experimental designs, and simulation-grounded proposals) under domain constraints, and as a systems-level capability whose value depends on how those candidates are validated, operationalized, and integrated into closed-loop discovery pipelines. We present a two-axis taxonomy organized by (i) generation modality and (ii) embodiment level, and use it to analyze seven core paradigms: knowledge-graph and literature-based discovery, symbolic regression and causal discovery, Bayesian experimental design, simulation-grounded scientific machine learning, representation learning and foundation models, diffusion-based and diversity-aware generation, and LLM-based hypothesis generation. We synthesize three recurring integration patterns, namely generator--verifier pipelines, latent-space optimization, and orchestrated closed-loop execution, and consolidate a six-dimensional evaluation framework covering efficiency, reliability, interpretability and idea quality, causal validity, discovery novelty, and mechanistic depth. The survey highlights the engineering interfaces that determine whether generative capability translates into dependable, auditable discovery pipelines.
- Keyword
- Idea generation, AI for Science, scientific discovery, closed-loop experimentation, scientific foundation models, evaluation framework.
- KSP Keywords
- Bayesian experimental design, Closed-loop, Domain Constraints, Generative capability, Idea quality, Integration patterns, Literature-based discovery, Molecular graphs, Scientific discovery, Scientific foundation, Space optimization
This work is distributed under the term of Creative Commons License (CCL)
(CC BY NC)
(CC BY NC)
