为什么在制药生产 AI 中，信任比智能更重要

来源：Dassault Systèmes 官方博客（BIOVIA 品牌专栏）

发布日期：2026-10-16

作者：Larry Fiegland、Jaya Tikhe（Dassault Systèmes）

原文：blog.3ds.com/brands/biovia/why-trust-matters-more-than-intelligence-in-pharmaceutical-manufacturing-ai/

Why Trust Matters More Than Intelligence in Pharmaceutical Manufacturing AI

想象一个 AI 模型，它比常规报警提前数小时标记出一次潜在的工艺偏差。预测结果看起来很有说服力。但工艺科学家能解释模型为什么发出这条警报吗？团队能判断这个信号是否有意义吗？在任何人采取行动之前，又需要什么证据？这些问题说明，AI 在制药生产中的落地不只是一个技术问题，而是一个科学确信与运营信任的问题。

Imagine an AI model that flags a potential process deviation hours before a conventional alarm would trigger. The prediction looks convincing. But can the process scientist explain why the model raised the alert? Can the team tell whether the signal is meaningful? And what evidence is needed before anyone acts? These questions show that AI adoption in pharmaceutical manufacturing is not just a technology problem. It is a problem of scientific confidence and operational trust.

图：3DS 官方博客题图，AI 驱动的生物工艺监测仪表盘，中央标记「Potential deviation detected」（检测到潜在偏差），右侧为四项信任支柱（图源：blog.3ds.com）

一、准确率只是起点

Accuracy is only the beginning

Pharmaceutical manufacturers generate enormous volumes of process data, and AI can help find the patterns in it. But a highly accurate prediction is not automatically suitable for a manufacturing decision. Relying on accuracy alone has three limits.

制药企业产生海量的工艺数据，AI 能帮助从中找出模式。但一个高准确率的预测并不自动适用于生产决策。只依赖准确率有三个局限。

Prediction is not explanation. A model may correctly flag an unusual batch without revealing the process relationships behind it. Knowing that something is different is not the same as knowing why.

预测不等于解释。模型可以正确标记出一条异常批次，却不揭示其背后的工艺关系。知道「有什么不一样」不等于知道「为什么不一样」。

Past performance is not a guarantee. A model is only as dependable as the conditions it was trained on. When raw materials, equipment, scale, or data collection change, its reliability can erode without anyone noticing.

历史表现不是保证。模型的可信度取决于它训练时所处的条件。当原材料、设备、规模或数据采集方式发生变化时，其可靠性可能在无人察觉的情况下衰减。

An anomaly is not automatically a problem. Normal variation, data-quality issues, and genuine process shifts can produce similar signals. Telling them apart takes scientific judgment.

异常不自动等于问题。正常波动、数据质量问题与真实的工艺漂移可能产生相似的信号。区分它们需要科学判断。

Accuracy measures one aspect of performance. It does not establish scientific validity or fitness for a particular manufacturing use.

准确率只衡量性能的一个侧面。它不能确立科学有效性，也不能证明对某一特定生产用途的适用性。

二、信任一条 AI 洞察意味着什么

What does it mean to trust an AI insight?

In practice, trust rests on four pillars:

在实践中，信任建立在四根支柱上：

Scientific interpretability (Explainable AI). Can the scientist see which variables or patterns drove the result, and does the explanation make sense for the process?

科学可解释性（可解释 AI）。科学家能否看到是哪些变量或模式驱动了这个结果？这个解释对这条工艺来说是否讲得通？

Data integrity and context. Is the data contextualized by batch, process stage, equipment, and operating conditions? Are missing values and changes in data sources understood? A sophisticated model built on poorly contextualized data is only sophisticated at being wrong.

数据完整性与情境化。数据是否按批次、工艺阶段、设备和运行条件完成了情境化？缺失值与数据源变更是否被理解？建立在糟糕情境化数据之上的精密模型，只是精于出错。

Reliability and governance. Has the model been evaluated for its intended use, with limitations, version, and conditions of applicability documented? Can the organization tell when performance deteriorates?

可靠性与治理。模型是否针对其预期用途做过评估？局限性、版本与适用条件是否有文档记录？组织能否察觉性能何时劣化？

Human judgment and accountability. Can scientists review the evidence, challenge a recommendation, and make a reasoned decision, with clear responsibility for the outcome?

人的判断与问责。科学家能否复核证据、质疑一条建议、做出有依据的决策，并对结果负有明确责任？

Trust doesn’t mean accepting an AI answer without question. It means having enough evidence to judge when that answer should, and should not, be used.

信任不是不加质疑地接受 AI 的答案，而是拥有足够的证据去判断这个答案什么时候该用、什么时候不该用。

三、从黑箱预测到工艺理解

From black-box predictions to process understanding

Consider a biologics batch. A model flags an unusual trajectory across glucose, lactate, viable cell density (VCD), feed rate, dissolved oxygen and pH, even though every parameter is within its operating limits. Univariate alarms stay silent. The model is seeing something in the combination.

考虑一个生物制品批次。模型标记出葡萄糖、乳酸、活细胞密度（VCD）、补料速率、溶氧与 pH 共同构成的一条异常轨迹，尽管每个参数都在各自的操作限值之内。单变量报警保持沉默。模型看到的是「组合」里的东西。

A scientist’s investigation might follow five steps:

科学家的调查可能遵循五个步骤：

1. Identify the deviation. Which variables contributed to the alert?

1. 定位偏差。哪些变量对这条警报有贡献？

2. Establish context. At what stage did the pattern emerge, and how does it compare with reference batches?

2. 建立情境。这个模式在哪个阶段出现？与参考批次相比如何？

3. Evaluate the evidence. Is it repeatable? Could it be measurement error, an operational difference, or a real process change?

3. 评估证据。它可重复吗？会不会是测量误差、操作差异，或真实的工艺变化？

4. Investigate explanations. What mechanisms could account for the behavior?

4. 调查解释。什么机制可以解释这一行为？

5. Decide on a response. Does the evidence call for further investigation, increased monitoring or action under established procedures?

5. 决定响应。证据是否需要进一步调查、加强监测，还是按既定程序采取行动？

图：制药生产线上的操作员正在设备旁检查（图源：3ds.com，BIOVIA Discoverant 产品页）

Notice what the model did and didn’t do. It found a pattern consistent with a potential issue. It did not prove a cause. Causal conclusions need additional evidence and rigorous analysis. This is where multivariate monitoring, reference-batch comparison and explainable machine learning support, rather than replace, scientific reasoning.

注意模型做了什么、没做什么。它发现了一个与潜在问题一致的模式，但没有证明因果。因果结论需要额外的证据与严格的分析。多变量监测、参考批次对比与可解释机器学习在这一点上是支持科学推理，而不是取代它。

四、信任必须被设计进工作流

Trust must be designed into the workflow

Not every AI application carries the same regulatory weight. An exploratory tool a scientist uses to investigate historical data is very different from a system supporting a GMP quality decision. That is the difference between generating an AI insight and using it in a regulated workflow.

不是每个 AI 应用都承担同样的监管权重。科学家用来调查历史数据的探索性工具，与支持 GMP 质量决策的系统完全不同。这就是「生成一条 AI 洞察」与「在受监管工作流中使用它」之间的区别。

Four considerations matter most:

最要紧的有四点：

Intended use. Define what the model does, which decisions it supports and what happens if it’s wrong.

预期用途。定义模型做什么、支持哪些决策、错了会怎样。

Validation and lifecycle management. Evaluate for the intended application, document limitations and reassess as conditions change.

验证与生命周期管理。针对预期应用做评估、记录局限性，并在条件变化时重新评估。

Traceability. Preserve the link between source data, model outputs, evidence and human decisions.

可追溯性。保存源数据、模型输出、证据与人的决策之间的链接。

Risk-based oversight. Match controls and human review to the risk, especially where product quality is affected.

基于风险的监督。让控制措施与人工复核同风险相匹配，尤其是在产品质量受影响的地方。

图：操作员在工艺控制面板前核对仪表读数（图源：3ds.com，BIOVIA Discoverant 产品页）

Trustworthy AI is a lifecycle responsibility spanning data, models, processes, people, and governance. It cannot be bolted on after development.

可信 AI 是贯穿数据、模型、工艺、人员与治理的生命周期责任。它无法在开发完成之后再加装上去。

五、下一个前沿：AI 辅助调查

The next frontier: AI-assisted investigation

Analytics is moving from static dashboards toward tools that detect patterns, explain contributing factors and ultimately assist investigations by letting scientists question contextualized data, test competing hypotheses and identify what evidence is still missing. AI agents may be the next step, orchestrating analytical workflows and retrieving process evidence. But they raise the stakes. An agent must not simply produce a convincing narrative. Its claims should be grounded in traceable evidence, its uncertainty visible and its actions constrained by the permissions and controls of its role. The goal is not to replace the process scientist. It is to help that scientist investigate faster and decide with more confidence.

分析技术正从静态仪表盘走向这样一类工具：检测模式、解释影响因素，并最终辅助调查，让科学家就情境化数据提问、检验相互竞争的假设、识别还缺什么证据。AI 智能体可能是下一步，负责编排分析工作流与调取工艺证据。但它们把赌注抬得更高。智能体不能只产出一个令人信服的叙事。它的主张必须落在可追溯的证据上，其不确定性应当可见，其行动应受其角色的权限与控制约束。目标不是取代工艺科学家，而是帮助这位科学家更快地调查、更有把握地决策。

六、目标是更好的决策，而不是更多的 AI

The goal is better decisions, not more AI

Predictive performance matters, but it is only one component of responsible deployment. Interpretability, contextualized data, appropriate validation and human accountability determine whether an insight can actually be used. The ultimate measure of AI in pharmaceutical manufacturing won’t be how sophisticated its algorithms become. It will be whether scientists can use its insights to understand their processes more deeply, investigate problems more effectively, and make decisions with greater confidence.

预测性能重要，但它只是负责任部署的一个组成部分。可解释性、情境化数据、恰当的验证与人的问责，决定一条洞察是否真的可用。衡量制药生产中 AI 的终极标准，不会是其算法变得多么复杂，而是科学家能否借助它的洞察更深入地理解自己的工艺、更有效地调查问题、更有信心地做出决策。

资料来源：Larry Fiegland, Jaya Tikhe, Why Trust Matters More Than Intelligence in Pharmaceutical Manufacturing AI, Dassault Systèmes 官方博客（BIOVIA），2026-10-16。配图取自 3DS 官方博客原文与 3ds.com 的 BIOVIA Discoverant 产品页。