Agentic Scientific Machine Learning for Neural Operators
- Question Elaboration
The research question focuses on the "cutting edge intersection of Agentic AI and Scientific Machine Learning (SciML)," specifically targeting the autonomous discovery of neural networks and neural operators for Partial Differential Equations (PDEs) and fluid simulations.
Key Terms & Scope:
Agentic AI: Autonomous systems capable of planning, reasoning, and tool-use to achieve long-term scientific goals without constant human intervention (Ali & Dornaika, 2025).
Neural Operators: Architectures like Fourier Neural Operators (FNO) or DeepONet that learn mappings between infinite-dimensional function spaces, allowing for mesh-independent PDE solving (Li et al., 2025; Wang et al., 2026).
Autonomous Discovery: The transition from manual hyperparameter tuning to "self-driving" research loops that iterate on model architectures and physical constraints (Ali & Dornaika, 2025).
PDEs & Fluid Simulations: High-dimensional, non-linear physical systems (e.g., Navier-Stokes) that serve as the primary benchmark for the discovered models.
- Key Concepts
Based on the elaboration, the following themes will structure the research:
Agentic Architectures for SciML: How multi-agent systems (MAS) replace human researchers in the "model-discovery-verification" loop.
Autonomous PDE Surrogation: The mechanics of LLM-driven discovery for physics-informed neural networks (PINNs) and operators.
Operator Learning & Fluid Dynamics: Modern neural operator benchmarks and their integration into autonomous workflows.
Closing the Loop: Literature retrieval, reasoning, and synthetic data generation as tools for agentic discovery.
- Web Search & Analysis Synthesis
Concept 1: Agentic Architectures for SciML
Current research identifies a shift from "passive" AI tools to "collaborative partners" (Ali & Dornaika, 2025). Systems like AgenticSciML employ over 10 specialized agents (proposers, critics, evaluators) to discover modeling strategies that outperform human-designed baselines by up to four orders of magnitude (Li et al., 2025).
Concept 2: Autonomous PDE Surrogation
The PINNsAgent framework exemplifies this trend, utilizing an LLM-based policy to explore a "Memory Tree" of hyperparameters (Wuwu et al., 2025). This eliminates reliance on expert heuristics by encoding PDE properties (linearity, dimension, boundary conditions) into feature vectors for similarity-guided exploration (Wuwu et al., 2025).
Concept 3: Operator Learning & Fluid Dynamics
Neural operators are uniquely suited for complex geometries in fluid dynamics. Recent benchmarks show that Vision Transformer (ViT)-based models with binary mask representations can enhance performance by 10% in intricate flow scenarios (Rabeh et al., 2025). Agentic loops can autonomously select these representations based on the problem's geometric complexity.
Checkit out here: https://github.com/moatasim-KT/SciMLx
or https://huggingface.co/spaces/hugging-science/SciMLx_Production
References
Ali, M. A., & Dornaika, F. (2025). Agentic AI: A Comprehensive Survey of Architectures, Applications, and Future Directions. arXiv. https://arxiv.org/pdf/2510.25445 Cited by: 69
Li, S., Chen, Y., Guo, Y., Huang, M., & Xiong, H. (2025). AgenticSciML: Collaborative Multi-Agent Systems for Emergent Discovery in Scientific Machine Learning. arXiv. https://arxiv.org/pdf/2511.07262 Cited by: 2
Wuwu, Q., Gao, C., Chen, T., Huang, Y., Zhang, Y., Wang, J., ... & Zhang, S. (2025). PINNsAgent: Automated PDE Surrogation with Large Language Models. arXiv. https://arxiv.org/pdf/2501.12053 Cited by: 9
Rabeh, A., Herron, E., et al. (2025). Benchmarking scientific machine-learning approaches for flow prediction around complex geometries. PMC. https://pmc.ncbi.nlm.nih.gov/articles/PMC12578797/
