OpenAIR iconOpenAIR
Open · AI · Research

A Trustworthy Closed-Loop
Agent Platform for Scientific Discovery

AI advances scientific discovery inside a complete research loop — humans stay in command, evidence stays verifiable

Validated in crystalline discoverythin-film PVD optimizationmore science domains · new domains arrive as skills

Problem Modeling→ Solution Design→ Execution & Validation→ Insight Generation ⟲ new computational results and experimental feedback re-enter any stage — a trustworthy closed loop
⭐ GitHub Repository 👀 Live Demo (no login) Up in 5 minutes
4-stage orchestrationstate management · checkpoint rollback
Contracted skillsregistry · composition · contract tests
Layered evidenceprovenance · versions · run records
Dry/wet loopmeasurement ingest · plan iteration
Demo

The trustworthy loop in 45 seconds#

From a natural-language start to four-stage progress, artifacts and the evidence chain

Full release and more: GitHub Release v0.1.0 · Direct link: pdl-openair.github.io/OpenAIR/en/#demo

Why OpenAIR

Four old research problems, four closed-loop answers#

OpenAIR models "how research gets done" as an orchestrable, verifiable, traceable four-stage process — not a hard-coded pipeline

PainLong cycles, fragmented tooling: literature, compute, experiments and writing each live in their own silos, with humans as the glue
OrchestrableOne four-stage process model; capabilities ship as skills, scheduled by a unified registry; long-horizon tasks run continuously
PainResults that don't reproduce: different person, different machine, different outcome — nobody can pinpoint the deviation
VerifiableSkills are defined by SKILL.md + I/O contracts, independently testable and replayable (replay mode reproduces the full pipeline with zero keys)
PainUntraceable evidence: which input, which version, which run produced this conclusion? No way to audit afterwards
TraceableEvery artifact carries a provenance chain — which skill, which inputs, which version — forming a trustworthy evidence chain
PainUnderused feedback: experimental results come back but rarely flow into the next round of design and decisions
Closed-loop iterationDry-lab (computational) and wet-lab (fabrication) feedback iterate in one loop, converging toward discovery
Validated Scenarios

Validated scenarios: materials science#

The platform is domain-agnostic; materials discovery and fabrication are the first fully validated scenarios — more domains arrive as skill packs

Materials Discovery

Crystalline Material Discovery

From target properties through candidate generation, evolutionary search and computational validation (VASP band gap / stability), producing a data-backed candidate list and analysis report.

mSUN 58.0%final search evaluation
candidates → evolutionary search → DFT validation → insight report
Open the full workbench snapshot →
Materials Fabrication

AlN Thin-Film PVD Optimization

Literature evidence structuring → fabrication-parameter recommendation → wet-lab fabrication → feedback-driven redesign, iterated over four rounds with rising piezoelectric performance.

d33 15.9 → 18.6four wet-lab feedback rounds
evidence → parameter recommendation → wet lab → feedback ×4
Watch the closed loop in action →
Mechanisms

Four mechanisms: from constrained execution to auditable evidence#

Meeting the three engineering challenges of taking research agents from demo to reusable application — hard-to-constrain execution · hard-to-close feedback · hard-to-audit process

🔁 Four-Stage Task Orchestration

problem_modeling → solution_design → execution_validation → insight_generation, with stage-state management and checkpoint rollback. L2 generic orchestration layered over L3 domain-atomic skills pushes structured I/O constraints and contract tests down to every capability call — keeping flexibility while raising controllability.

🧩 Contracted Skill Management

200+ multi-domain skills organized by standardized frontmatter, directory structure and schemas — with registry, retrieval, composition, versioning and contract testing. A new domain is a new skill pack (no platform code changes), and replay reproduces an entire run with zero keys.

🧾 Layered Evidence Tracing

Skill provenance, licenses, stage mappings and migration records are linked to task executions and artifacts; the trace chain extends from sources and versions to computation parameters, experimental data, runtime environments and conclusion-verification grounds — supporting audit and reproduction.

🧪 Dry/Wet-Lab Feedback Linkage

Adapts to the cloud computing and cloud-lab platforms so measured data flows directly into the next round of plan revision and orchestration, with feedback-driven stage rollback and iteration; automatic feedback-driven policy learning is on the roadmap.

Quick Start

Three ways to begin#

A · Zero install
Live demoOpen the full workbench snapshot in a browser — no login, read-only.
B · Local, zero keys
openair CLIpipx install openair-cli && openair up — a local instance with replay in about 10 minutes.
C · Full capability
Manual Docker deploygit clone + docker compose up; configure a model gateway and external accounts for live mode.
# Option B: openair CLI
pipx install openair-cli
openair up            # clone → configure (random secrets + admin) → start → open browser

# Option C: Docker
git clone https://github.com/pdl-openair/OpenAIR.git
cd OpenAIR && cp .env.example .env
docker compose up -d --build
# Workbench http://localhost:8080 · API http://localhost:8000/health