科研 Skill

自驱动实验室闭环材料发现

high-throughput-autonomous-labs

Covers combinatorial synthesis, robotic experimentation. Use for autonomous or robotics-driven experimental campaigns, a less standardized domain.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

High-Throughput and Autonomous Materials Discovery

Overview

Autonomous/self-driving labs close the loop between synthesis, characterization, and decision-making — an algorithm (typically Bayesian optimization or a related active-learning method) selects the next experiment based on prior results, without waiting for a human to manually plan each iteration. This is a genuinely different workflow from every other skill in this repo, which assume a human plans each characterization/synthesis step; it's included because it's real and growing, but the tooling here is less standardized than characterization or DFT conventions.

When to use this skill

Activate when the request involves:

  • High-throughput synthesis, combinatorial screening, autonomous lab, self-driving lab, active learning for materials discovery, Bayesian optimization (applied to synthesis/experiments)
  • Terms: acquisition function, exploration-exploitation, closed-loop experimentation, design of experiments for screening
  • "Plan a high-throughput screening campaign," "set up Bayesian optimization for synthesis parameters"

Core usage

Bayesian optimization for synthesis parameter selection

from skopt import gp_minimize
from skopt.space import Real

def objective(params):
    """Wraps the actual synthesis + characterization pipeline — in a real
    autonomous lab, this triggers hardware and waits for a measured result;
    here it's the interface point where this skill hands off to
    synthesis-processing and the relevant characterization skill."""
    temperature, precursor_ratio = params
    measured_property = run_synthesis_and_characterize(temperature, precursor_ratio)  # placeholder for the real pipeline
    return -measured_property  # gp_minimize minimizes; negate if maximizing the target property

search_space = [Real(400, 800, name="temperature_C"), Real(0.5, 2.0, name="precursor_ratio")]
result = gp_minimize(objective, search_space, n_calls=30, random_state=42)
best_params = result.x

Design-of-experiments screening (before optimization — establishing the response surface)

from scipy.stats import qmc

def latin_hypercube_screening_design(n_samples, parameter_bounds):
    """Space-filling design for initial screening — better coverage of the
    parameter space than a naive grid for a given experiment budget, standard
    practice before committing to a more targeted optimization campaign."""
    sampler = qmc.LatinHypercube(d=len(parameter_bounds))
    unit_samples = sampler.random(n=n_samples)
    lower_bounds, upper_bounds = zip(*parameter_bounds)
    return qmc.scale(unit_samples, lower_bounds, upper_bounds)

Validation & Pitfalls

Canonical references: Shahriari et al. (2016), "Taking the human out of the loop: A review of Bayesian optimization," Proceedings of the IEEE, for the general methodology; MacLeod et al. (2020), "Self-driving laboratory for accelerated discovery of thin-film materials," Science Advances, for a representative autonomous-lab implementation.

  • Bayesian optimization's efficiency gains depend on the underlying response surface being reasonably smooth/well-behaved — a genuinely discontinuous or highly multimodal property landscape can mislead the surrogate model and the optimizer toward a local rather than global optimum. Don't assume convergence to a good result without checking whether the explored space and found optimum make physical sense.
  • The measurement/characterization step feeding the optimization loop has its own noise and systematic error (see the relevant characterization skill's Validation & Pitfalls) — an optimization loop built on noisy or biased measurements can converge confidently to a wrong answer. The optimization algorithm doesn't know the measurement is unreliable; validate measurement quality independently, not just optimization convergence behavior.
  • This field's tooling and conventions are less standardized than the rest of materials characterization — a workflow that works well for one lab's specific hardware/software stack often doesn't transfer directly to another's. Treat the code here as illustrating the underlying optimization concepts, not a drop-in pipeline; real autonomous-lab integration requires hardware/software-specific engineering this skill doesn't cover.
  • A successful autonomous campaign optimizes the specific objective function defined — a poorly chosen objective (e.g. optimizing one property while ignoring a constraint like stability or cost) can produce a "successful" result that's practically useless. Define the objective function (including any constraints, as penalty terms or explicit constraint handling) carefully before running a campaign, not after seeing where it converges.
  • Batch/parallel experimentation (common in robotic labs, running many experiments simultaneously rather than one at a time) requires optimization strategies designed for batch acquisition — a purely sequential Bayesian optimization strategy applied naively to batch selection can pick redundant (too-similar) experiments within a batch; use a batch-aware acquisition strategy when the lab setup actually runs experiments in parallel.

示例提示词

帮我设计一个贝叶斯优化引导的机器人合成闭环实验方案。

所属插件

材料表征、性能与材料信息学面向实验与数据驱动的材料研究:XRD/电镜/谱学/热分析/力学测试等表征,合成与加工,电化学与半导体表征,相场、CALPHAD 与连续介质力学,材料数据库、特征工程、贝叶斯优化与主动学习,选材与生命周期评价。

安装该插件会一并获得同领域的 42 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

自驱动实验室闭环材料发现能做什么?

Covers combinatorial synthesis, robotic experimentation. Use for autonomous or robotics-driven experimental campaigns, a less standardized domain。

如何安装自驱动实验室闭环材料发现?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「自驱动实验室闭环材料发现」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

自驱动实验室闭环材料发现需要哪些权限?

该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。

自驱动实验室闭环材料发现经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。