AgenticSciML
Multi-agent system in which more than ten specialized LLM agents debate, retrieve, and evolutionarily refine scientific-machine-learning architectures and training procedures.
| Affiliation | Brown University, Division of Applied Mathematics (Karniadakis lab) |
| First introduced | 2025-11 (arXiv:2511.07262) |
| Lifecycle stages | Hypothesis (new SciML modeling strategies) + Experiment design (architectures, loss formulations, training procedures) |
| Autonomy level | Semi-autonomous (human supplies problem statement, requirements, evaluation criteria, and approves the evaluation contract; agents then run iterative tree expansion) |
| Domain focus | Scientific machine learning — physics-informed neural networks (PINNs), neural operators, PDE-constrained learning, inverse problems |
| Availability | Code on request — repository announced “available on GitHub upon publication” in the preprint |
Approach
Three-phase pipeline. (1) The human user provides Problem.md, Requirements.md, Evaluation.md, and an optional Data_config.json. (2) A multi-modal data-analyst agent performs exploratory data analysis on the training set and writes a data_analysis.md for downstream text-only agents; an evaluator agent then writes a formal evaluation contract (evaluate.py and guidelines.md) that the human approves. (3) Specialized agents — knowledge retrievers, proposers, critics, engineers, debuggers, and a result-analyst — collaboratively expand a tree of candidate solutions through ensemble-guided evolutionary search. The best-scoring solution is always retained for mutation (exploitation), while additional parents are selected by majority vote of a diverse ensemble of selector agents (exploration). A persistent methodological memory stores prior solutions, ablations, and error analyses; structured debate forces proposers and critics to justify and revise modeling decisions.
Validation
Benchmarked on PDE-constrained learning, operator learning, and inverse problems. The authors compare against single-agent LLM systems, automated PINN frameworks, and NAS / AutoML baselines.
Notable results
- Reports up to four orders of magnitude error reduction over single-agent and human-designed baselines (sometimes 1000×) on physics-informed and operator-learning tasks.
- Discovers SciML strategies that do not appear in the curated knowledge base, including adaptive domain-decomposed PINNs for multi-scale PDEs, physics-informed operator-learning architectures with constraint-conditioned branches, and dynamically weighted loss schedules derived from residual-flow structure.
Primary paper
Other references
None yet.
Code
Not released at preprint time; the paper states code and experiment results will be available on GitHub upon publication.