orchid
A modular Python framework that lets a local LLM run a Find → Reason → Simulate → Validate loop against materials-science and chemistry data, with grounding against the Materials Project and PubChem at every step.
Origin
Most LLM-agent demonstrations are tool-calling loops without a domain. orchid is the result of asking what the same loop looks like when it has to ground in something with real validation surfaces — not “did the model answer plausibly”, but “does the prediction match a measurement”.
The chosen domain is materials science and chemistry: well-defined inputs, computable physics, validated reference data, and a real bar to clear. The framework is structured to make the agent's reasoning auditable rather than just impressive.
How it is structured
Librarian
Vector store (ChromaDB) over scientific documents, materials, and chemicals, with an optional graph store (Neo4j) for relationships. The agent queries this before reasoning, so its starting context is grounded rather than hallucinated.
Chemist & Physicist
RDKit for molecular validation with PubChem cross-checking; Pymatgen for crystal stability; physics simulation wrappers for stress, thermal, and fatigue analysis, with optional FEM. Every prediction carries an uncertainty alongside it.
Optimiser
Multi-objective (NSGA-II-style) Pareto search over candidate materials for a target specification. Selects “best for this constraint set” rather than “best on a single metric” — the shape of real engineering trade-offs.
Validation
Materials Project for DFT-calculated material properties and PubChem for validated chemistry. Built-in experimental benchmarks from ASM Handbook, MatWeb, and CRC Handbook. RMSE, MAE, and 95% confidence intervals reported alongside every comparison.
The autonomous loop
At the centre is a self-directed researcher that picks its own next task, runs the full pipeline, and accumulates discoveries. The same component works as a one-shot session or a long-running daemon.
The researcher identifies gaps in coverage, picks the next material or hypothesis to explore, runs the full Find → Reason → Simulate → Store cycle, validates against real data, and persists discoveries so subsequent queries get smarter.
Why it is public
The Find → Reason → Simulate → Validate scaffolding is reusable for any domain with a similar shape: a knowledge base, a reasoning model, a simulator, and a validation source. Engineering design, computational biology, applied chemistry, and a handful of other fields fit the same mould.
Releasing orchid lets that pattern be lifted out and adapted rather than rebuilt. The materials-science backend can be swapped for any other discovery problem with the same loop.
orchid on GitHub
Source, documentation, and the autonomous-research entry points.