Project

PDEScientist

An agentic PDE-discovery system for proposal, evaluation, and refinement.

PDEScientist is the next stage after direct symbolic prediction. The core idea is to let a model propose candidate PDEs while evaluator tools handle target derivatives, feature construction, coefficient fitting, and residual scoring.

The project is motivated by a practical gap in scientific discovery. A one-shot model can make a plausible symbolic guess, but difficult cases often require inspecting residuals, trying alternative terms, fitting coefficients, and revising the hypothesis. PDEScientist studies that iterative loop.

What the project studies

  • Multi-turn equation proposal and revision.
  • Tool feedback for coefficient fitting and residual diagnostics.
  • Evaluation contracts that prevent the model from defining its own target.
  • Trace-based analysis of why a proposed equation succeeds or fails.

Design principles

  • The evaluator owns the target derivative and scoring logic.
  • The model proposes symbolic equations, but it should not be able to redefine the task.
  • Tool feedback should be structured enough to support revision, not just pass/fail reporting.
  • Each run should leave an inspectable trace of proposals, fitted coefficients, residuals, and final selection.

Why it is different from a single model

Scientific discovery often requires iteration. A system should be able to make a proposal, test it, understand the mismatch, and use that feedback in the next attempt.

Back to homepage