A breeding workbench that carries a real experiment from field record to defensible conclusion — five stages, one evidence chain, and a version history that now runs from V1.0 to V13.0.
Spend time in a field and one thing becomes obvious: breeding teams rarely lack
an algorithm. What they lack is the continuity between the things around the
algorithm — the number a material is given in the field, the images taken of it,
the table that comes out of cleaning, and the conclusion a paper eventually
reports. When any one of those loses its correspondence to the others, no model
downstream can repair it.
So the platform grew around a single requirement: a researcher does one real
thing in one place, and can say where every result came from and how to check
it. Internally that is called an evidence chain, and most of the engineering
went into the parts of it that demo badly and matter most.
The five stages
The platform's working loop is five verbs, with the cross-cutting concerns —
permissions, parameter records, versioning, quality review, audit — running
underneath all of them rather than living in a module of their own.
采 / Collect. Field images, manual trait records, experiment designs,
genotype matrices and environment data, imported in batches with their own
metadata intact: material number, plot, time, device, operator, conditions. For
multi-device scenes — drones, smart cameras — an explicable import and
validation protocol comes first; the hardware itself is supplied by the scene's
owner or settled by agreement, and third-party capture equipment is not counted
as something the team built.
选 / Index. Sample–experiment–trait–environment–file connected into a
searchable catalogue, with field mapping, batch filtering, sample-consistency
checks and an exception log. Same name for different things, different names for
the same thing, repeated measurements, one sample recorded at several time
points: the original record is never silently overwritten. The mapping rule,
the conflict list and the trace of a human confirmation are kept beside it.
洗 / Clean. Missing-rate statistics, duplicate checks, field-type and unit
validation, outlier annotation, the intersection of genotype and phenotype
samples, and the train/predict split — each pass leaving a log and a rollback
snapshot. Outliers are annotated by default rather than quietly deleted, and the
imputation method with its scope of use goes into the report, so two people's
cleaning habits cannot produce two contradictory conclusions.
算 / Compute. Standard statistics, trait recognition, genetic association
and predictive modelling wrapped as configurable tasks. Every task emits a model
version, random seed, data hash, parameters, dependency environment, runtime,
error log and result files, so the same input can be re-run to within a stated
tolerance. Parallel scheduling serves throughput, but it comes with resource
quotas, retries and consistency checks — no throughput that cannot be reviewed.
用 / Use. Interactive charts, candidate material or candidate locus lists,
confidence intervals and applicability boundaries, and shareable, traceable
reports. Downstream effects like "preferred material" or "a shorter breeding
cycle" get their own field validation: a model score or a pipeline's runtime
does not stand in for a breeding result.
From the last chart back to the data
Original data is not overwritten — that is one of the floor rules. Every
transformation produces a traceable derived version tagged with its input
source, cleaning rules, execution environment, algorithm version and owner, and
a manual change must carry its reason. The thing to be able to do is this: take
the final chart in a report and locate the dataset and workflow configuration
behind it.
The interface is built to be usable, but not at the price of hiding the
assumptions. The professional judgements and the risks of use belong where a
reader can see them.
Iteration: V1.0 to V13.0
The route ran 需求管理 (V1.0) → 工作台雏形 (V3.0) → 流程编排 (V7.0) →
场景化集成 (V13.0). Worth stating plainly: thirteen iterations is a versioning
and development-activity figure, not thirteen separately accepted commercial
releases. From here, every version closes the loop — defects in the last one,
user feedback, the change list, test results, release — so that iteration stops
being a presentation and becomes a software process.
Where this actually stands
This is the section I least want to blur. Three levels have to stay apart:
| Level | State |
| --- | --- |
| A demonstrable prototype | Data interface, workflow orchestration, tool scheduling, parallel-task UI |
| Tried in real research tasks | Pilot scenarios with teams at national key laboratories |
| Expected at the productisation stage | Permission isolation, data audit, stable deployment, standardised support |
Acceptance is not "clicking produces a result" either. It has to answer: does a
failure say clearly what went wrong; can a task be interrupted and resumed; does
the same input return a result within a stated numerical tolerance; are large
uploads, weak networks and permission conflicts handled properly; is there a
complete record of deletion, backup and restore. And for a researcher who does
not write code, the question is whether they can go from import through quality
control and analysis to a report without a developer operating it for them.
My role
Project lead. The positioning, the development route, the platform architecture
and the five-part division are mine to propose and push; version iteration, the
intellectual-property record, pilot communication and documentation sit on this
side too. A good deal of the algorithm implementation, interface development and
field work belongs to other members of the team — what this page describes is my
share of it, and it does not turn a team's output into one person's.
Intellectual property
Three software copyrights are registered under the platform: 智能集成机器学习建模与
可视化分析平台 (2025SR1800232), 自适应可视多平台主成分分析软件 (2024SR1708999) and
面向多需求的生物数据处理软件 (2024SR1711525). Patents and papers are listed on
their own pages, item by item, with name, number, rights holder and current
legal status.
A software registration, a patent application or grant, and a paper under review
or published are three different states. This site does not merge them into one
sentence about having "several achievements".
The research programmes it sits inside
The work belongs to two national programmes, and the platform exists to serve
them:
面向生物育种的科研智能方法研究 — 2025ZD0121801
生物育种国家科技重大专项 — 2023ZD04076
What is deliberately not on this page
No screenshots of the competition deck, no business-plan layout, no headline
metrics quoted from slides. The plan is confidential and its layout — not its
content — is what must not circulate, so what appears here is the verifiable
part: programme numbers, registrations, pilot scenarios, the iteration path.
About the numbers. The roadshow material carried claims like "about 100,000
maize plants", "130× faster" and "35% more accurate". Those need re-checking
against their original reports first, and stating crop, sample size, device,
input resolution, annotation rules, baseline algorithm, hardware, repetitions
and error interval before they belong anywhere public; two models that did not
run under the same conditions do not get a single ratio between them. Until they
are stated that way, they stay off this page — which is the same standard this
site holds anyone else's numbers to.