Spectroscopic methods play an instrumental role in the implementation of the U.S. Food and Drug Administration outlined process analytical technology for biopharmaceutical manufacturing. Industrial spectroscopic calibration models are typically developed in an offline setting using traditional methods, such as partial least squares and principal component regression. Apart from the limiting performances of these conventional models under time‐varying operating conditions, these methods require access to extensive historical data, which are seldom available in biopharmaceutical manufacturing. In this article, we propose a novel spatiotemporal just‐in‐time learning (ST‐JITL) based spectroscopic model calibration platform for automatic training and maintenance of calibration models using routine batch data. The proposed ST‐JITL framework uses Gaussian processes (GPs) for local model calibration. A GP model not only exhibits superior performance over traditional methods but also provides credibility intervals around the model predictions. The efficacy of the ST‐JITL based model calibration platform is demonstrated in predicting the critical performance parameters of an industrial cell culture process.

中文翻译：

---

使用时空实时高斯过程实时监控细胞培养过程的光谱模型

光谱方法在美国食品和药物管理局概述的生物制药生产过程分析技术的实施中发挥了重要作用。工业光谱校准模型通常是使用传统方法（例如，偏最小二乘和主成分回归）在离线环境下开发的。除了这些常规模型在时变操作条件下的局限性之外，这些方法还需要访问广泛的历史数据，而这些历史数据在生物制药生产中很少可用。在本文中，我们提出了一种新颖的基于时空实时学习（ST-JITL）的光谱模型校准平台，用于使用常规批处理数据自动训练和维护校准模型。拟议的ST‐JITL框架使用高斯过程（GPs）进行局部模型校准。GP模型不仅表现出优于传统方法的性能，而且提供了围绕模型预测的可信区间。基于ST‐JITL的模型校准平台的功效在预测工业细胞培养过程的关键性能参数中得到了证明。