规划合成路线、设计实验、审计实验方案、分析化合物和反应，并在工作需要模拟时进入计算化学。
在进入实验台前规划合成路线、设计实验并审计实验方案。
化学工作区
把本页作为化学产品界面的入口：合成规划、实验与实验方案设计、工具支持的化学答案、笔记本和计算。更深入的页面会解释每个区域，而不让本页承载所有细节。
从目标、论文、专利或反应草图开始。维思纳帮助把它转化为可行路线和下一步实验问题。
草拟有来源依据的步骤，暴露假设，并根据化学、物理、危险性和操作约束审计计划。
为化合物和反应生成实用事实包：性质、危险性、谱图、产物、配平、相容性和证据缺口。
使用科学模拟完成化学计算、绘图、数据分析、PySCF、RDKit 和可复现笔记本工件。
当轻量分析不够时，进入范围受控的分子建模工作流和部分 Rowan 驱动的高级计算。
当维思纳需要分子量、沸点、GHS 编码、反应产物或建模结果时，它会调用合适工具并保持证据可见。当工具无法验证某件事时，答案会明确说明。
分子量来自 RDKit；沸点来自 Thermo 库；GHS 危险性来自 PubChem。答案中的每个数值都会引用生成它的库和函数，便于你验证或自行重跑。
当工具没有某个化合物的数据时，它会返回“数据覆盖缺口”信息，而不会说“未检测到危险”。随后智能体会回退到自身化学知识，并把那部分标注为推理，而不是已验证结果。
一次实验方案审计可能调用收率、原子平衡、相容性、沸点和安全工具。智能体会选择运行哪些工具、按什么顺序运行，并读取输出生成最终答案。你只需描述一次问题。
产品界面
智能体可以从实际化学目标开始，然后根据任务需要路由到本地工具、笔记本、实验方案工作流或外部计算。
先起草，再证明式检查。
实验方案生成器会区分有来源依据的步骤和假设。实验方案审计会形式化步骤，生成待验证项，并把发现、缺口、假设和来源记录为证书。
范围受控的定量化学。
化学计量、理论收率、溶液配制、稀释、pH、简单平衡、蒸气压和理想气体压力。缺失常数或模糊物种会产生验证错误，而不是猜测。
用于智能体推理的证据包。
化学智能解析化合物身份、性质、危险性、溶解性、谱图和相似性。反应智能增加反应解释、预测、配平、相容性、能量学、动力学和明确数据缺口。
以笔记本为先的建模，需要时由 Rowan 提供支持。
AI 会为描述符、pKa、构象、DFT、对接、MD 和其他分子建模任务创建可见笔记本。部分高级工作流由 Vicena Compute 中的 Rowan 支持，并受额度限制。
你可以完成什么
下面每个场景都是真实化学任务。查看智能体会做什么、返回什么，以及它可以路由到哪个更深入的产品区域。
你从论文中提取合成路线，或自己设计一条路线。你即将投入时间、试剂和玻璃器皿。维思纳会读取实验方案、检查计算、标记不相容项，并在约 30 秒内引用它发现的内容。
你会看到什么
逐步审计：理论最大收率与主张收率、原子平衡、溶剂沸点与回流温度、每种试剂的 GHS 危险性，以及相容性担忧（漂白剂 + 氨、HF + 玻璃、氯化溶剂 + 聚碳酸酯）。每个数字都引用来源库。无法验证的内容会标记为数据覆盖缺口，而不是被掩盖。
它节省什么
半小时手工 SDS 审阅，一次组会中“等等，这个收率真的可能吗？”的尴尬，以及本会浪费在不可行方案上的试剂。
论文、补充信息和专利常把方法埋在密集文字中。维思纳会提取温度、浓度、时长和试剂等变量，与来源主张交叉核对，并标记不一致或不完整内容。
你会看到什么
结构化实验方案，每个参数都有标签，每个值都有置信度和来源页/段落。可直接粘贴到实验记录本。
它节省什么
节省你因为补充信息和正文数字不同而反复阅读方法部分三遍的一个小时。
对于 Suzuki、Buchwald、Heck、酰胺偶联和 Fischer 酯化等专利常见反应，维思纳可以调用训练好的反应模型。对于 SN2、酸碱和简单酯化等教材反应，它可以使用确定性规则。你会得到预测产物、所用方法和清晰可靠性边界。
你会看到什么
预测产物的 SMILES 和结构、生成它的方法，以及预测可靠和不可靠的边界。
它节省什么
减少已知反应上的心智消耗，把思考留给未知反应。
粘贴一个 SMILES，获得 LogP、MW、HBD、HBA、TPSA、可旋转键、Lipinski 通过/失败、Veber 通过/失败，以及与你选择的参考化合物的 Tanimoto 相似性。
你会看到什么
完整描述符表和基于规则的评估，加上参考集相似性评分和骨架分析。
它节省什么
省去五个 SwissADME 标签页和手动指纹比较，同时保持方法一致和输出结构化。
你想要描述符、pKa、构象、HOMO-LUMO 能隙、优化几何、偶极矩或振动光谱。用自然语言提出即可。维思纳会创建笔记本，选择合适的本地或 Rowan 驱动工作流，在预算限制内运行并返回数字。
你会看到什么
一个 Jupyter 笔记本，包含输入、使用 Rowan 时的工作流 UUID、带正确单位的结果、表格、图表、CSV 文件，以及可用时与参考值的比较。笔记本会保留：你可以修改、扩展，或在方法部分引用。
它节省什么
节省为单个建模问题准备输入文件、轮询、提取描述符和绘图的设置时间。
应用说明
了解维思纳如何把化学目标转化为结构化合成规划、工具支持检查和可供决策的研究工件。
下面是示例提示。点击即可运行。
Example prompts
五个类别，每类三个提示，从快速查询到多步骤工作流。点击任意提示即可用它打开新聊天。
GHS 危险性、相容性检查和实验方案级审计。
带安全交叉检查的溶剂选择。原子经济性和收率分析。
身份、结构、物理性质、危险性、溶解性、谱图、描述符和相似性。
正向产物、逆合成、配平检查、相容性、能量学和动力学。
PySCF 笔记本，加上通过 Vicena Compute 运行的 Rowan 驱动分子建模。
Some examples run on the free tier; notebook and Rowan-powered workflows require Plus or Ultra.
完整工作流
下面每个工作流都是化学家每天解决的真实问题。点击任意示例即可在新聊天中运行。
在进入实验台前捕捉不可能收率、不安全组合和隐性危险。
Retrosynthetic routes with literature backing and predicted yields.
Properties, predicted spectra, and functional groups in one pass.
Lipinski + Veber assessment with safety and solvent compatibility.
Ranked solvent compatibility cross-referenced with GHS hazards.
DFT, Hartree-Fock, or MP2 calculations written and executed in your notebook.
实验方案审计
Given a protocol, the agent breaks it into steps and calls each of the checks below. They return numerical results with citations, or "coverage gap" if the underlying database doesn't have data. The LLM reads the tool output and writes the final summary; it does not invent numbers the tools didn't return.
Computes theoretical maximum from reactant mass and molecular weights via RDKit. A claimed yield over 100% is flagged as a conservation of mass violation.
Checks that every atom in the products came from somewhere in the reactants. Catches "atoms from nowhere" errors in proposed mechanisms.
Curated rules + SMARTS-based classification (HF + glass, alkali + protic, chlorinated + polycarbonate, bleach + ammonia) with resolver that handles compound aliases and qualifiers.
Exact values from the Thermo library (DIPPR correlations, 70,000 compounds), cited by source so you know what the AI knows.
Live GHS classification from PubChem: signal word, hazard codes, precautionary statements. Refreshed per query, never cached stale.
When a check has no rule or data for your compound, it says so explicitly. Never "no hazards detected". Always "I cannot verify, use your judgment."
Who typically uses these tools.
Who it's for
For professionals and students. For bench chemists and theorists. The barrier Vicena lowers is the same in both cases: the friction between reading about chemistry and actually doing it.
Catch impossible yields and unsafe combinations before running a protocol.
Protocol Audit + Safety SummaryEvaluate drug-likeness, find analogs, plan synthesis routes.
Retrosynthesis + Descriptors + SimilarityReview student protocols at scale, with cited safety reasoning.
Protocol Audit + GHS + CompatibilityPredict spectra, pick solvents, understand unknown molecules.
Spectroscopy + Solubility + Functional GroupsRun molecular modeling workflows in visible notebooks without wiring every SDK call by hand.
Vicena Compute + Rowan + JupyterAll the tools
Every tool the agent can use for chemistry work. Click any tool for the full reference.
Quantitative chemistry calculations
Compound identity, properties, hazards, and fingerprints
Reaction prediction, feasibility, compatibility, energetics, and kinetics
Properties and hazards lookup
Neural network reaction prediction
Work backwards from target molecule
Textbook reaction predictions
SMARTS substructure matching
Drug-likeness and Lipinski analysis
Tanimoto fingerprint comparison
GHS hazard information
Name, SMILES, and CAS conversion
Advanced molecular modeling and notebook workflows
Open-source libraries under the hood.
Libraries used
Open-source chemistry libraries and public databases. The specific library or database behind each answer is cited in the output.
Industry-standard cheminformatics toolkit used at Pfizer, Novartis, and Merck.
NIH's open chemistry database: 116 million compounds with structures, properties, and safety data.
70,000 compounds with validated thermodynamic properties (DIPPR correlations).
Academic quantum chemistry package for Hartree-Fock, DFT, and post-HF calculations.
Selected advanced molecular modeling workflows are powered by Rowan through Vicena Compute, gateway routing, and computation credits.
Physical chemistry library for stoichiometry, balancing, and kinetics.
Neural network trained on 1.2 million patent reactions (IBM / EPFL).
Common questions
Yes. The theoretical yield calculator uses reactant mass and molecular weights from RDKit to compute the 100% theoretical maximum. If your claimed yield exceeds that, the tool flags a conservation of mass violation. Ask: "Is 6.2 g of product from 5 g of starting material physically possible for SMILES X → Y?"
A curated rule set plus SMARTS-based structural classification. Catches HF + glass, alkali + protic solvent, bleach + ammonia (chloramine gas), chlorinated solvent + polycarbonate, and about 20 other common incompatibilities. When no rule matches, the tool returns a coverage gap message instead of implying the pair is safe.
ReactionT5v2, a sequence-to-sequence neural network trained on 1.2 million patent reactions (USPTO). Published accuracy is ~93% on patent-style organic reactions, lower on organometallics and unusual chemistries. Simple textbook reactions (esterification, SN2, hydrolysis) use deterministic SMARTS rules instead.
Yes, using the retrosynthesis mode of ReactionT5v2. Give it a target SMILES and it proposes disconnections and starting materials. Best for drug-like organic molecules; less reliable for very large molecules, polymers, or inorganic targets.
Those tools are curated databases (SciFinder, Reaxys) or single-purpose reaction predictors (IBM RXN). Vicena integrates multiple open-source components (RDKit, PubChem, Thermo, ChemPy, PySCF, ReactionT5v2) through an LLM that orchestrates them. It does not replace the full literature search of SciFinder or Reaxys; it covers property lookup, prediction, audit, and quantum chemistry tasks.
Pulled live from PubChem per query, not cached. You get the current signal word, hazard statements (H-codes), precautionary statements (P-codes), and pictogram codes for any compound PubChem indexes.
Vicena supports local notebook workflows such as Hartree-Fock, DFT, and MP2 via PySCF, plus selected molecular modeling workflows powered by Rowan through Vicena Compute. Budget caps and computation credits apply before expensive submissions.
Yes. The molecular descriptors tool computes MW, LogP (Wildman-Crippen), HBD, HBA, and TPSA via RDKit and evaluates against Lipinski's Rule of Five and Veber's rules for oral bioavailability.
Functional-group-based IR, ¹H/¹³C NMR, and MS fragment prediction using a SMARTS-matched spectral database. Not quantum-mechanical; use PySCF via the notebook for first-principles IR.
It means the tool looked for the compound or reaction and could not find data for it. Rather than returning an empty result (which looks like "safe" or "no hazards"), the tool explicitly says the database does not cover this case and the LLM falls back to its own chemistry knowledge, labeled as reasoning rather than as a verified result.
From the co-founder
We're trying to reduce the friction between reading chemistry and doing chemistry. Professionals and students should be able to extract ideas from papers, design experiments, audit protocols, and run computational chemistry without learning every tool one by one.
The goal is to let researchers focus on the scientific question while Vicena handles the mechanics: which tool to call, what assumptions matter, what evidence is available, and where the result stops being verified.
Think of it as Deep Work for chemistry: describe the objective in natural language, let the AI pick the right tools, and keep the evidence visible from the first idea through the executed result.
This is a small step in our journey toward AI for real science, helping students and professionals do science more enjoyably, with less friction.
Free to start. Paste a protocol and ask for an audit, look up a compound, or plan a reaction.
