01
f(X) + ε ~ N(0, Σ)Statistical modeling
Understand the data, before the model learns.
Sample traits · variable relations · domain structure
Statistics × AI,
intelligence you can trust.
Beyond generating answers.
Every exploration moves toward grounded judgment.
How is a judgment
made trustworthy?
With limited samples, complex constraints and high reliability demands,
statistical checking runs through every AI exploration.
02
P(Y | do(X))Causal inference
Moving together does not mean driving each other.
Identification conditions · influence mechanisms · experiment design
03
μ(x) ± zα·σ(x)Uncertainty quantification
Give the judgment, and its boundaries.
Result intervals · model risk · valid boundaries
04
min L(x) s.t. g(x)≤0Optimization & decisions
The last step: make the plan survive reality.
Multi-objective trade-offs · real constraints · feasible options
MatNavis
An AI agent platform for materials R&D
More than intelligence:
evidence you can rely on
Materials data × domain algorithms × large models
Projects as the unit of work, connecting data, models, tasks and results.
Showcase01:48
01/ UNDERSTANDKnowledge Retrieval & Dialogue
Every answer cited, every lead grounded.
Start research sessions in natural language. The planner orchestrates retrieval across public papers and private project docs with inline citations.
向材料智能体交办科研任务
用自然语言描述研究问题，智能体自动拆解目标、组织公私域检索与模型验证
梳理太赫兹超构表面吸波结构的核心设计变量与实验验证数据，核对 S11/S21 边界依据。
私有数据1.4 GB48 组超构表面 X/Y 已解析
私有知识3 个空间128 篇太赫兹论文与实验记录
私有模型4 项可执行FR-CVAE / 频谱代理模型就绪
近期研究任务6 项运行中最新：宽频反射谱结构反演
直接对话 · 材料文献检索与机理探讨已从总览发起会话 ↗
梳理太赫兹超构表面吸波结构的核心设计变量与实验验证数据，核对 S11/S21 边界依据。
📖knowledge.public-papers混合检索命中 2 篇太赫兹超构表面高相关论文 (Score 0.94)✓
🔒knowledge.private-lab关联课题组《超构表面参数扫描记录.csv》#14–#19 实测数据✓
太赫兹超构表面吸波与相移核心参数核验结论✓ 公私域交叉核验通过
S11 反射机理
在 0.96 THz 处由开口谐振环电容耦合产生反射极小值 -18.4 dB，阻抗匹配带宽覆盖 0.88–1.12 THz[1]。
S21 相移与加工边界
0.8–1.2 THz 内实现 267° 连续透射相移（透射幅度 > 0.82），受限于光刻工艺最小线宽 w ≥ 12 μm 与间隙 g ≥ 10 μm[1][2]。
公域论文 [1]相似度 0.94 · P4 Fig.3b
Terahertz metasurface inverse design with aligned X/Y samples核实 C4 对称微元在 0.8–1.2 THz 的 267° 相位覆盖规律🔒 私域实测 [2]扫描记录 #14–#19
超构表面参数扫描记录.csv · 暗室实测复核验证线宽 w < 12 μm 时边缘蚀刻偏差导致 S21 衰减 > 1.8 dB02/ PREDICT & DESIGN9-Stage Inverse Design & Forward Prediction
Forward from structure, inverse from target.
Align design variables X, observed responses Y, and target constraints across a 9-stage workflow to infer spectra and search candidate topologies.
九阶段逆向设计 · 候选拓扑生成与频谱核对阶段 8 / 9 · 在线推理就绪
✓ 已对齐项目样本与目标约束槽位（5 / 5）MOCK-frcvae-design & response
参考样本设计 X12组 8×8 网格 ✓
参考样本响应 Y1001 频点双通道 ✓
几何约束X 边界线宽 ≥ 12 μm ✓
频段约束Y 范围0.8–1.2 THz ✓
优化目标目标响应连续透射相移 ✓
✦FR-CVAE 逆向生成与代理模型联合推理潜空间采样 12 组候选结构，已完成 1001 频点相移与透射谱核验✓
candidate-0199.2% 最优candidate-0298.1%candidate-0396.7%0.8–1.2 THz 频段响应拟合相移覆盖 269°
预测曲线 (candidate-01)目标相移曲线
启动 FR-CVAE 逆向搜索，筛选满足最小线宽 ≥ 12 μm 的高匹配微元拓扑并核对频段响应。
03/ DECIDE & AUDITActive Learning DoE & Managed Execution
Guided next experiments, auditable end to end.
Analyze feature distributions and error bounds to recommend Bayesian active-learning samples, guarded by human approval gates and trace logs.
增量优化与受管执行 · 贝叶斯主动学习补点Iteration v2 → v3
上一轮研究结果 · Iteration v2当前样本集：48 组超构表面样本 · 代理模型交叉验证
0.80–1.04 THz 平坦区MAE 0.019 · 预测达标 ✓
1.05–1.15 THz 相位突变区⚠️ MAE 0.068 · 存在高不确定性盲区
◈贝叶斯主动学习 · skill.bayesian-doe基于期望改进准则（EI）锁定突变区 3 组高信息量补点方案✓
特征盲区靶向采样分布盲区误差 -35%
既有样本 (48)推荐补点 (+3)
doe-next-01g = 14.2 μm · p = 118 μmEI 0.94 · -38%doe-next-02g = 16.5 μm · p = 124 μmEI 0.89 · -31%doe-next-03g = 12.8 μm · p = 112 μmEI 0.84 · -26%审批门控 · run-0241版本留痕：v1 (36组) → v2 (48组) → v3 (+3组补点)
已锁定输入样本快照与超参数配置，人工复核通过后启动下一轮增量电磁仿真。
针对 1.05–1.15 THz 相位突变盲区执行贝叶斯主动学习，推荐下一批补点方案并提交审批。
Every conclusion traces back to its evidence · From question to reusable Artifact Traceable Evidence · Auditable Process · Explicit Boundaries · Expert Review
Methods leave the model
and enter the field.
Starting from real questions,
closing the decision loop in R&D, production, operations and management.
Put a limited experiment budget into better-grounded exploration.
- Experiment design & sample-size planning
- Key mechanism variables identified
- Process and formulation optimization
- Confidence assessment for R&D conclusions
R&D scenarioApproach
Grounded experiment and formulation candidates
Problem input
Experiment records / composition / performance targets
Extract patterns→Small-sample modeling→Iterate on feedback
R&DWhere should the next experiment start?
Put a limited experiment budget into better-grounded exploration.
- Experiment design & sample-size planning
- Key mechanism variables identified
- Process and formulation optimization
- Confidence assessment for R&D conclusions
R&D scenarioApproach
Grounded experiment and formulation candidates
Problem input
Experiment records / composition / performance targets
Extract patterns→Small-sample modeling→Iterate on feedback
ProductionWhat actually causes the quality drift?
Process, quality and equipment information are joined to trace the key variables.
- Process parameter sensitivity
- Root-cause modeling for quality & yield
- Equipment anomaly and fault alerts
- Side-by-side process comparison
Production scenarioApproach
Process options that fit production
Problem input
Process parameters / equipment state / quality records
Sensitivity analysis→Root-cause modeling→Option comparison
OperationsDemand shifts — how do resources keep up?
Demand dynamics are understood and operating strategies compared under resource and cost constraints.
- Dynamic demand forecasting
- Resource allocation under constraints
- Operations and marketing evaluation
- Cost and return attribution
Operations scenarioApproach
Comparable resource and operating options
Problem input
Demand series / resource constraints / operating records
Forecast demand→Evaluate strategy→Allocate resources
ManagementHow far does one decision reach?
Metric attribution and scenario simulation make the reasoning open to review.
- Sensitivity attribution across KPIs
- In-flight and ex-post policy evaluation
- Macro and business scenario simulation
- Full-chain decision audit
Management scenarioApproach
Explainable, traceable management evidence
Problem input
Business KPIs / policy conditions / assumptions
Metric attribution→Scenario simulation→Process audit
News & insight
The technology moves; we are on site with it.
Notice of 2026 X-AGI Conference
More than a dozen tracks — AI4Science, agents, LLM foundations and evaluation among them — from 16 to 18 October at the Beijing Friendship Hotel.
Read the article
UDI Participates in the 4th Joint Conference on Statistics and Data Science and the IMS China Biennial Meeting
At the year’s gathering of statistics and data science, we talked with academic and industry partners about taking “statistics × AI” from method to practice.
Why Vertical Industries Need Statistics × AI
Under limited data and high reliability demands, statistical modeling decides how far AI can go — our read on what vertical industries actually need.
StatIDE: Embedding Statistical Expertise in the Development Environment
Statistical consulting expertise usually stays in the expert’s head. StatIDE turns it into a workbench the whole team can reuse.
Different intelligences,
one pursuit of truth.
Unified Data Intelligence Technology Co., Ltd.
Unified Data Intelligence
Through the deep integration of statistics, data science and AI, building a stronger and more dependable generation of AI.
From generating answers to dependable judgment and scientific decisions.
That is how we connect technology and industry.
Partners
With universities, research institutes and industry partners,
bringing trustworthy intelligence to real problems.
Join us
Whether you lean towards algorithms, data, engineering or product — if applying statistics, data science and AI to real problems is what interests you, we would like to hear from you.
Apply nowWhat we look for
- A solid foundation in statistics, data science, artificial intelligence or large models
- Strong learning ability and self-drive, and the discipline to finish the work you take on
- Ready to communicate and to review what worked and what did not
- Familiar with at least one programming language
- Course, research, competition or personal projects are a plus