Manual crystallization trials have historically posed significant challenges, demanding substantial expertise for process development and often offering unpredictable outcomes. This study addresses these difficulties by introducing an automated system that alleviates the need for manual iterations and intuitive deductions. The system leverages machine learning algorithms capable of learning from high-quality data to discern patterns and recommend optimal actions for subsequent runs. The automation process commences with a direct chord length (DCL) control system, generating system-specific training data via universal crystallization rules. After that, the automation process will progress into a machine learning iteration loop using adaptive neuro-fuzzy inference system (ANFIS) models. In this iteration loop, multiple models will be built (with accumulative historical data) and deployed to the crystallization process until predefined exit criteria are met or a maximum of five iterative cycles are reached. Results from the two campaigns are presented. It is evident that the automated crystallization platform with machine learning’s ability can confidently explore the operational space, proposing credible processing conditions that yield desirable process outcomes.

中文翻译：

---

ANFIS 驱动的机器学习自动化平台，用于冷却结晶工艺开发

手动结晶试验历来提出了重大挑战，需要大量的工艺开发专业知识，并且常常会产生不可预测的结果。本研究通过引入自动化系统来解决这些困难，该系统减少了手动迭代和直观推论的需要。该系统利用能够从高质量数据中学习的机器学习算法来识别模式并为后续运行推荐最佳操作。自动化过程从直接弦长 (DCL) 控制系统开始，通过通用结晶规则生成系统特定的训练数据。之后，自动化过程将进入使用自适应神经模糊推理系统（ANFIS）模型的机器学习迭代循环。在此迭代循环中，将构建多个模型（使用累积的历史数据）并将其部署到结晶过程，直到满足预定义的退出标准或达到最多五个迭代周期。介绍了这两项活动的结果。显然，具有机器学习能力的自动化结晶平台可以自信地探索操作空间，提出可靠的加工条件，从而产生理想的工艺结果。