17,564 public projects17,564 个公开项目
The Project Universe项目宇宙
Every public ISEF finalist project from 2014–2026, arranged by what it is actually about.2014–2026 全部公开的 ISEF 决赛项目，按研究内容排布。
A category tells you which room a project was judged in. It does not tell you what the project was about. This page arranges every public abstract by its own text, so the neighbourhoods are research subjects rather than administrative labels — and so you can see, before you commit to a topic, how crowded the ground around it already is.类别只说明一个项目在哪个房间里被评审，说明不了它究竟在研究什么。这一页把每一份公开摘要按它自己的文字排布，于是聚集起来的是研究主题而不是行政标签——你也就能在定题之前，先看清你想去的那块地方已经有多挤。
17,564
public projects个公开项目
10,632
content links条内容关系
22
tracks个赛道
2014–2026
seasons covered届全覆盖
Pick a track选一个赛道
Twenty-two tracks二十二个赛道
Open any track to see its projects as a map. The bar is how much of the corpus that track accounts for; the line on the right is how its intake moved across the seasons.点开任意一个赛道，把它的项目看成一张图。条形是该赛道在全库中占的分量，右侧折线是它历届入围数量的走势。
Loading…正在载入…
Before you read it读之前先知道
What this map is, and what it is not这张图是什么，不是什么
- What puts two projects near each other什么让两个项目靠在一起
- Their own words. Every project is placed by the text of its public title and abstract, nothing else — not its awards, not its country, not its school.是它们自己的文字。每个项目的位置只由它公开的标题和摘要决定，不看它得了什么奖，也不看它来自哪个国家、哪所学校。
- Distance is not a score距离不是分数
- Flattening this much text onto a page loses information. Two projects sitting close are usually related, but how far apart two dots are does not measure how different the work is, and a big cluster is not a hot field.把这么多文字压平到一张纸上，一定会丢信息。挨得近的通常确实相关，但两个点之间的远近不代表两项研究差多少，一片大的聚集也不代表那个方向就热。
- A line is a resemblance, not a relationship连线是相像，不是关系
- Lines are drawn where two abstracts read alike enough to pass a fixed threshold. They are not citations, not collaborations, and not one project building on another. Projects that resemble nothing are left unconnected rather than tied to something for the sake of the picture.两份摘要写得够像、越过固定门槛，才会连一条线。它不是引用，不是合作，也不是谁在谁的基础上做的。跟谁都不像的项目就让它孤立着，不会为了画面好看硬给它接上。
- Nobody is named不涉及任何人的姓名
- Names, schools and every other identifying field are excluded from the computation and from this page. Where a student happened to cite their own earlier work by name inside their abstract, that name is masked.姓名、学校以及任何身份字段都不参与计算，也不出现在这一页上。个别学生在自己的摘要里按姓名引用了此前的工作，这类姓名已被遮蔽。
The full method and its numbers完整方法与口径
- Positions位置
- Each track is fitted on its own: one- and two-word TF-IDF over title (weighted three times) and abstract, with common and generic research wording removed; truncated to 64 dimensions by SVD; laid out in two dimensions by t-SNE on cosine distance with a fixed random seed, so the layout is reproducible and filtering never reshuffles the points.每个赛道单独拟合：对标题（三倍权重）与摘要做单词和双词组 TF-IDF，去掉常用词与通用科研措辞；用 SVD 截断到 64 维；再以余弦距离做 t-SNE 投影到二维，随机种子固定，因此布局可复现，筛选也不会把点重新打乱。
- Links连线
- Computed independently of the positions, from the full TF-IDF vectors: a link exists where one project is among the other’s four nearest by cosine similarity and that similarity is at least 0.16. Average similarity along a link runs 0.22–0.25 by track, against at most 0.06 for all project pairs within the same track.独立于位置计算，用的是完整 TF-IDF 向量：当一个项目在另一个的余弦相似度前四名内、且相似度不低于 0.16 时成立。各赛道连线的平均相似度在 0.22–0.25 之间，而同赛道全体项目两两之间最高只有 0.06。
- Directions方向分组
- KMeans over the same vectors, named by their highest-weight terms. These are exploratory text groupings, not official subcategories — the official ones are on the track pages. The grouping is computed in 64 dimensions while the picture has two, so the groups overlap on screen: across all tracks only about 64% of projects sit closest to their own group’s centre.KMeans 在同一批向量上求得，用权重最高的词命名。这些是探索性的文本分组，不是官方子类别——官方子类别在赛道页上。分组在 64 维空间求得而画面只有二维，所以分组在图上是互相交叠的：全部赛道合计只有约 64% 的项目离自己所属分组的中心最近。
- Coverage覆盖范围
- 17,564 public finalist projects, 2014–2026, from the official abstract database.17,564 个公开的决赛入围项目，覆盖 2014–2026，取自官方摘要库。
Official subcategories and award records by track →各赛道的官方子类别与获奖记录 →
The full data analysis →完整数据分析 →