Bgolearn Playground
小白入门材料贝叶斯优化A beginner path to Bayesian optimization for materials
从“试很多次”变成“聪明地试下一次”：用少量实验数据训练代理模型，让不确定性和采集函数帮你推荐最值得做的材料候选。 Move from trying everything to choosing the next experiment wisely: train a surrogate model from limited data, then use uncertainty and acquisition functions to recommend high-value material candidates.
先把流程讲成人话Plain-language walkthrough first
它不是替你凭空发明材料，而是把你已经做过的实验整理成一张“会标出把握程度的地图”。地图越清楚，下一次实验就越像有方向的移动，而不是盲猜。 It does not invent materials from nothing. It turns your completed experiments into a map that also shows confidence. The clearer the map becomes, the more directed the next move feels.
点击右侧任一步，看数据在贝叶斯优化循环里怎么移动。 Click any step on the right to see how data moves through the loop.
高斯过程会看你已经测过的点，估计其他位置可能好不好。蓝线像“助手的猜测”，金色不确定性像“这里我还不太确定”。The Gaussian process looks at measured points and estimates untested regions. The blue curve is the assistant's guess; the gold band says where it is still unsure.
采集函数会给候选排序。它既看预测性能，也看不确定性：有时去最高点附近，有时去还没摸清的区域。The acquisition function ranks candidates. It considers predicted performance and uncertainty: sometimes it goes near the best region, sometimes into blind spots.
推荐只是“值得一试”的建议。执行实验后，系统会显示 GP 原本怎么预测、实际结果是多少、误差有多大。A recommendation is only a worthwhile suggestion. After the experiment, the page shows the GP prediction, the measured value, and the prediction error.
只有点击“加入数据”后，新实验才会进入训练集。下一轮重新拟合 GP，地图更新，推荐也会跟着改变。Only after clicking Add data does the measurement enter the training set. The next GP fit updates the map, so the recommendation changes too.
材料实验常常昂贵、缓慢、样本少。贝叶斯优化把已有实验变成一张“带不确定性的地图”，再决定下一步去哪里探索。 Materials experiments are often costly, slow, and data-limited. Bayesian optimization turns existing measurements into an uncertainty-aware map, then decides where to explore next.
把成分、工艺、温度、结构特征等转成模型能读懂的变量。Convert composition, process, temperature, and structure descriptors into model-readable variables.
用代理模型预测性能，同时估计“这里我有多不确定”。Use a surrogate model to predict performance and estimate uncertainty at the same time.
采集函数按不同策略给候选排序：有的偏向提升性能，有的偏向降低未知区域。An acquisition function ranks candidates by strategy: some favor performance gain, while others reduce unknown regions.
新实验结果回到数据表，模型更新，推荐会越来越聚焦。The new measurement goes back into the table, the model updates, and recommendations become more focused.
你是一间材料实验室的产品负责人。每一次点击代表真实实验室的一轮：高斯过程根据已有样本拟合性能地图，采集函数给候选打分，然后你花掉 1 次预算测量一个材料。目标是在预算耗尽前完成足够轮次的主动学习，并找到性能超过任务阈值的候选。 You are the product lead of a materials lab. Each click is one lab round: a Gaussian process fits a performance map from measured samples, an acquisition function scores candidates, and you spend one budget unit to measure a material. Your goal is to complete enough active-learning rounds and beat the mission target before the budget runs out.
把贝叶斯优化拆开看，其实就是模型、信心和决策规则。At beginner level, Bayesian optimization is model, confidence, and a decision rule.
本游戏里的拟合函数是高斯过程回归。它用 RBF 核表达“相近材料可能性能相近”，每轮实验后重新计算后验均值和不确定性。This game uses Gaussian process regression. Its RBF kernel encodes that nearby materials may have similar properties, then recomputes posterior mean and uncertainty after every experiment.
模型不仅给出预测值，还告诉你哪些区域“没把握”。高不确定区域可能藏着惊喜。The model gives both prediction and confidence. Regions with high uncertainty may hide useful surprises.
采集函数不是只有一种。它把模型的预测值、不确定性、当前最好结果和策略偏好转成候选排序。分数最高的位置，才会成为下一轮实验建议。There is not only one acquisition function. It turns prediction, uncertainty, current best value, and strategy preference into a candidate ranking. The top-ranked candidate becomes the next experiment.
直接奖励高预测和高不确定性。k 越大，越敢探索未知区域。Rewards high prediction and uncertainty. Larger k explores unknown regions more.
关心“比当前最好结果还能提升多少”的期望值，常用于性能最大化。Looks at the expected gain over the current best result, common for maximization.
关心“超过当前最好结果的概率”，更像是在找成功率高的位置。Looks at the probability of beating the current best, favoring likely wins.
更重视信息价值：这次实验能让我们对全局最优了解多少。Focuses more on information value: how much this experiment teaches about the global optimum.
Bgolearn 是面向材料发现的 Python 贝叶斯全局优化框架，支持回归、分类、主动学习、虚拟筛选和多种采集函数。 Bgolearn is a Python Bayesian global optimization framework for materials discovery, supporting regression, classification, active learning, virtual screening, and multiple acquisition functions.
pip install Bgolearn
查看接口、示例和使用说明。Read APIs, examples, and usage notes.
Bgolearn: a unified Bayesian optimization framework for accelerating materials discovery.Bgolearn: a unified Bayesian optimization framework for accelerating materials discovery.
从论文了解 Bgolearn 的整体框架、采集函数和材料发现应用。Read the paper for Bgolearn's framework, acquisition functions, and materials-discovery use cases.
python bgolearn_ui.py
然后打开 127.0.0.1:8787。Then open 127.0.0.1:8787.
轻量级表格机器学习工具包，可作为桌面应用或浏览器应用使用，支持回归、分类、聚类、可视化、预测和结果导出。A lightweight toolkit for tabular machine learning. It can be used as a desktop app or opened in a browser, and supports regression, classification, clustering, visualization, prediction, and result export.
轻量级科研绘图工作台，可从表格数据创建论文风格图、显示晶体结构、模拟 XRD、排版多面板图并导出最终图片。A lightweight scientific figure workspace. It helps users create publication-style figures from spreadsheet data, display crystal structures, simulate XRD patterns, arrange multi-panel figures, and export final images without writing plotting code.
如果 Bgolearn 支持了你的研究，请引用以下论文。If Bgolearn supports your research, please cite the paper below.
@article{Cao2026Bgolearn,
author = {Bin Cao and Jie Xiong and Jiaxuan Ma and Yuan Tian and Yirui Hu and Mengwei He and Longhan Zhang and Jiayu Wang and Jian Hui and Li Liu and Dezhen Xue and Turab Lookman and Jun Wang and Tong-Yi Zhang},
title = {Bgolearn: a unified Bayesian optimization framework for accelerating materials discovery},
journal = {npj Computational Materials},
year = {2026},
doi = {10.1038/s41524-026-02226-3},
issn = {2057-3960},
publisher = {Springer Nature},
url = {https://doi.org/10.1038/s41524-026-02226-3}
}