ChemAgent

Autonomous chemistry/chemical-engineering research agent. Plans, reasons, and solves ChemE problems using LLM tool-use over molecular, thermodynamic, computational, and literature tools.

Backend: claude -p — uses your Claude Code auth. No separate API key required.

v0.2 benchmark: 228/230 (99.1%), 22/22 numerical problems correct across 10 ChemE subdomains.

v0.3 (in progress): added polymer chemistry (RAFT kinetics), peptide / AMP descriptors, structured trace logging, and RAG over a polymer-chemistry corpus — 29 benchmark problems across 17 categories, 104 unit tests, 96.9% on the v0.3 baseline run.

Why

Most LLM agents are built by CS engineers against software tasks. This one is built by a chemical engineering student to tackle domain problems — unit operations, kinetics, thermodynamics, separations — the kind of work a process engineer or research chemist does. It's positioned at the intersection of AI agent building and chemical engineering, a combination that's scarce in 2026.

What it does

Given a problem like:

Design a CSTR for aspirin production via salicylic acid + acetic anhydride with rate r = k·[SA]·[AA], k = 0.001 L/(mol·s). Feed [SA]_0 = 2.0 mol/L, [AA]_0 = 2.2 mol/L (10% excess). Target 80% conversion of SA. Plant must produce 100 kg/day of aspirin (MW 180). Compute the required CSTR volume.

ChemAgent:

1. Plans a solution path (ReAct loop, JSON actions)
2. Calls tools — molecular lookup (RDKit), vapor pressure (Antoine), Python execution with scipy, arxiv literature search
3. Returns a reasoned answer with equations, units, and final value
4. Is scored by an LLM-judge against a ground-truth benchmark

Stack

• claude -p (ReAct JSON actions) — agent brain
• RDKit — molecular properties, SMILES parsing
• Antoine equation — vapor pressure for 6 common solvents
• Python sandbox — arbitrary math with math, numpy, scipy
• arxiv API — literature search, no key needed
• Fluids module — Colebrook/Swamee-Jain, three-reservoir networks, Crane TP-410 fittings
• Polymer module — RAFT kinetics (Arrhenius rate constants for 6 monomers, 3 initiators, 4 RAFT agents) with Müller dispersity, inverse target-DP design
• Peptide / AMP module — net charge (Henderson-Hasselbalch), Kyte-Doolittle hydrophobicity, Eisenberg hydrophobic moment μH, Chou-Fasman helix propensity — for SNAPP / antimicrobial peptide design
• Trace logging — every agent run writes a JSONL trace (traces/<run_id>/<problem_id>.jsonl) with timestamps, tool calls, latencies, and final answers. Replayable and analysable.
• RAG (BM25) — pure-Python BM25 over a polymer-chemistry corpus (corpus/). No ML dependencies. Currently covers RAFT mechanism + kinetics, AMP design, SNAPPs, polymer dispersity. Easily extended by dropping more .txt / .md files into the corpus.
• Pytest — 104 unit tests on tools
• LLM-judge eval harness — 29 problems across 17 ChemE subdomains

Setup

git clone https://github.com/jincinga24-hue/chem-agent
cd chem-agent
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Prerequisite: Claude Code installed and authenticated.

Run

# Single question
python examples/run_agent.py "Calculate vapor pressure of ethanol at 50 C"

# Full benchmark
python evals/run_eval.py

# Unit tests
pytest tests/

Architecture

src/
  agent.py        ReAct tool-use loop
  claude_cli.py   subprocess wrapper for claude -p
  prompts.py      system + judge prompts (numerical and qualitative rubrics)
  tools/
    molecular.py   RDKit wrappers (MW, logP, SMILES)
    thermo.py      Antoine vapor pressure, ideal gas
    python_exec.py sandboxed Python (math, numpy, scipy)
    literature.py  arxiv search (Atom API)
    fluids.py      pipe roughness, Reynolds, friction factors, three-reservoir solver
    polymer.py     RAFT polymerization kinetics + inverse target-DP design
    peptide.py     net charge, hydrophobic moment, helix propensity, AMP-likeness
    knowledge.py   BM25 RAG over corpus/ — keyword-grounded retrieval
  tracing.py     structured JSONL trace logger
corpus/             plain-text knowledge corpus (RAFT, AMPs, SNAPPs, dispersity)
evals/
  problems.json   29 benchmark problems + ground truth
  judge.py        LLM-judge scoring
  run_eval.py     runs agent on all problems with per-problem traces
traces/             JSONL traces, one file per problem per run
tests/
  test_tools.py     unit tests for molecular/thermo/python_exec/literature
  test_fluids.py    unit tests for fluids module
  test_polymer.py   unit tests for RAFT kinetics
  test_peptide.py   unit tests for peptide / AMP descriptors
  test_tracing.py   unit tests for trace logging
  test_knowledge.py unit tests for RAG / BM25

Benchmark: 29 problems across 17 ChemE subdomains

Category	Problems	Example
Molecular	1	MW of aspirin via RDKit
Thermodynamics	1	Antoine vapor pressure of ethanol at 50°C
Non-ideal thermo	1	Wilson activity coefficient for ethanol-water
Reactor design	3	CSTR, PFR (1st order), batch (2nd order)
Separation	2	Fenske min stages, Rachford-Rice VLE flash
Heat transfer	4	LMTD, NTU-effectiveness, composite wall, steam condenser
Bioreactor	1	Monod chemostat steady state
Mass transfer	1	Kremser absorption stages
Fluid flow	2	Ergun packed bed, pipe friction pump power
Fluid mechanics	2	Torricelli three-jet tank, three-reservoir junction (Colebrook)
Mass balance	1	Two-stream NaCl mixer
Unit conversion	1	SCFM methane → kg/h
Design	1	Multi-step aspirin CSTR sizing
Optimization	3	Wilson parameter fit, kinetic order regression, Underwood root
Literature	1	Qualitative arxiv search
Polymer	2	RAFT MMA Mn prediction; inverse target-DP design
Peptide / AMP	2	Melittin net charge at pH 7.4; α-helical hydrophobic moment

Roadmap

• v0.1 — single-step tool-use, 5 problems eval, claude -p backend
• v0.2 — 23 problems, 10 subdomains, scipy sandbox, arxiv tool, qualitative judge
• v0.3a — fluids (Colebrook, three-reservoir), polymer (RAFT kinetics), peptide / AMP descriptors
• v0.3b — structured trace logging (JSONL per problem) + RAG over polymer-chemistry corpus (BM25)
• v0.3c — extend corpus to full Odian/Moad chapters + Perry's Handbook excerpts; switch BM25 to embedding-based retrieval
• v0.4 — LangGraph refactor for parallel tool use, polymer informatics (polyBERT / property prediction)
• v0.5 — DWSIM integration for real process simulation
• v0.6 — lab-in-the-loop: agent proposes + runs experiments

Known limitations

• ReAct loop assumes Claude emits well-formed JSON. Multi-line Python code strings require json.loads(..., strict=False) — handled but is a known text-based tool-use failure mode. Native tool-use API (Anthropic SDK) would avoid it.
• Benchmark problems are textbook-style. Real research problems would need more ambiguity tolerance and missing-data handling.
• No persistent memory — each run is independent.
• Sandbox is not a security sandbox; it restricts imports but trusts the LLM.

License

MIT