GRAFT-ATHENA

Self-improving agentic framework that accumulates methodological experience across scientific problems and autonomously expands its own action space for the evolutionary discovery of numerical algorithms.

   
Affiliation Brown University, Division of Applied Mathematics (Karniadakis group)
First introduced 2026-05 (arXiv:2605.11117)
Lifecycle stages Multi-stage (problem formalization + method selection + implementation + advisor critique + cross-problem memory)
Autonomy level Semi-autonomous — user supplies a free-form problem; the system formalizes it, samples methods from its action tree, executes, scores outcomes, and accumulates experience
Domain focus Scientific computing — PDE solvers (Nektar++, Trixi.jl, PIML), dissipative particle dynamics (LAMMPS), and scientific machine learning
Availability Code on request — authors state the repository will be made publicly available on GitHub upon acceptance

Approach

GRAFT (Graph Reduction to Adaptive Factored Trees) projects directed acyclic graphs of solver attributes and cross-rules into factored decision trees with a deterministic embedding into a metric space. Each method is a single path through an action tree T_A and each problem a single path through a companion problem tree T_P; a persistent memory D records every solved instance with its observables and reward. This changes the policy footprint from exponential to linear in decision chains while certifying the factorization as an I-map of the policy. Multiple specialized teams coordinate on this substrate: Expansion and Construction teams ingest solver documentation to grow the action graph; a Formalization team audits equations, boundary data, reductions, and identifiability before encoding a problem as a fingerprint on T_P; a Strategy team samples methods from T_A conditioned on the fingerprint; an Implementation team realizes the choice as runnable code; and an Advisor scores the outcome, revises failed choices, or extends the action tree with new nodes mid-trial. Nearest-neighbor retrieval from D over fingerprint similarity supplies a reward-calibrated prior for each new problem.

Validation

On four canonical physics-informed machine-learning benchmarks, GRAFT-ATHENA reaches near-machine-precision accuracy, surpassing human and prior agentic baselines including the predecessor ATHENA. On production solvers it reconstructed Mach-10 hypersonic flow over the Apollo Command Module from a 1968 NASA report (Nektar++/Trixi.jl) and recovered shear-thinning red-blood-cell rheology in dissipative particle dynamics (LAMMPS). The viscous Burgers case demonstrates transfer: the prior retrieves a Reynolds-number continuation from neighboring high-Re runs and improves on the earlier agentic trace even at moderate Re.

Notable results

  • Autonomously proposed regularization constraints for ill-posed inverse problems and persisted them as new nodes in the action graph.
  • Designed a spectral PINN with exponential convergence — a new numerical method emerging from agentic exploration.
  • Demonstrates monotonically accumulating methodological memory across problems (every iteration, success or failure, commits to D).

Primary paper

Toscano, Chai, and Karniadakis, “GRAFT-ATHENA: Self-Improving Agentic Teams for Autonomous Discovery and Evolutionary Numerical Algorithms,” arXiv:2605.11117.

Other references

None yet.

Code

Authors commit to a GitHub release upon acceptance; no public URL yet.