| name | pievo |
| description | Autonomous scientific discovery via principle evolution. Trigger when the user wants to optimize experimental parameters, discover material design principles, or run multi-round automated experiments. Requires pievo.enabled=true in config AND scientific MCP tools (zeopp, raspa, mofstructure, mofchecker) to be available.
|
| pinned | true |
| metadata | {"miqi":{"emoji":"🔬","requires":{"mcp":"[Truncated]"}}} |
PiEvo — Principle-Evolvable Experiment Selection
When to use
- User asks to optimize parameters ("find the best MOF for CO2 capture")
- User wants to discover scientific principles from experimental data
- User wants to automate a multi-round experimental campaign
- User mentions "贝叶斯优化", "实验设计", "自动发现", "原则进化"
Prerequisites
pievo.enabled must be true: check with miqi config get tools.pievo.enabled- At least one scientific MCP tool must be connected (zeopp-backend, raspa-mcp, etc.)
- If dependencies are missing, tell user to run:
pip install miqi[pievo]
Workflow
Step 1: Understand the task
Extract from user input:
- Parameter space (what can be varied?)
- Optimization target (what metric to maximize/minimize?)
- Constraints (any hard limits?)
- Available tools (which MCP scientific tools are relevant?)
Step 2: Launch discovery session
pievo_run(
task="<natural language task description including parameters, target, constraints>",
tools=["<mcp_tool_name_1>", "<mcp_tool_name_2>"],
max_rounds=20,
warm_up_rounds=5
)
Step 3: Monitor progress
pievo_status("<session_id>")
Step 4: Report results
When the session completes, summarize for the user:
- Discovered principles and their belief scores
- Best parameters found
- Number of rounds executed
- Convergence trend
Important Notes
- PiEvo uses 3 LLM calls per round (Principle, Hypothesis, Experiment phases)
- The GP model needs warm_up_rounds of data before IDS activates
- If a tool call fails, the round is recorded with outcome=0.0 and PiEvo continues
- Sessions run asynchronously — use pievo_status to monitor, not polling in a loop
