Scholarlink

<p>The use of artificial intelligence (AI) is growing in pharmaceutical manufacturing, but regulatory assimilation has been slow due to difficulties in interpretability, uncertainty calibration and sustainability assessment. This work offers a quantitative framework for assessing the deployment of trustworthy AI by combining multi-objective composite trust index (CTI). The framework combines model predictive performance (R², RMSE), uncertainty measures (predictive variance and Expected Calibration Error), interpretability stability as well as sustainability metrics like Process Mass Intensity (PMI). With 120 industrial reaction batches, four AI paradigms (explainable AI (XAI), human-in-the-loop HITL, hybrid physics-informed AI, and uncertainty quantification UQ) were compared. The robust prediction performance of hybrid models was superior, with the highest R² and lowest RMSE among the methods considered. For example, we did see it helped with calibration when we used UQ to ensure that ECE was minimized. The human-in-the-loop technique also helped with greater consistency, thus enhancing robustness. Dimensions were normalized to generate a Composite Trust Index (CTI). This allowed us to be able to visualize that trade-off between the accuracy of long-term predictions versus how sustainable it could be. It is notable that the performance of PMI agents has a direct relationship with a large CTI: hence trust can be linked to the algorithm and greener process conditions. From the results, we know that Hybrid-UQ is the most reliable and sustainable among them with statistical significance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Sustainable Pathways to Trustworthy AI in Pharmaceutical Manufacturing: A Comparative Analysis of Explainable AI (XAI), Human‑in‑the‑Loop Systems, Hybrid AI, and Uncertainty Quantification Techniques

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月31日

查看全文
  • Nima Alizadeh Raef,
  • Siamak Pedrammehr,
  • Fatemeh Hamedi

摘要

The use of artificial intelligence (AI) is growing in pharmaceutical manufacturing, but regulatory assimilation has been slow due to difficulties in interpretability, uncertainty calibration and sustainability assessment. This work offers a quantitative framework for assessing the deployment of trustworthy AI by combining multi-objective composite trust index (CTI). The framework combines model predictive performance (R², RMSE), uncertainty measures (predictive variance and Expected Calibration Error), interpretability stability as well as sustainability metrics like Process Mass Intensity (PMI). With 120 industrial reaction batches, four AI paradigms (explainable AI (XAI), human-in-the-loop HITL, hybrid physics-informed AI, and uncertainty quantification UQ) were compared. The robust prediction performance of hybrid models was superior, with the highest R² and lowest RMSE among the methods considered. For example, we did see it helped with calibration when we used UQ to ensure that ECE was minimized. The human-in-the-loop technique also helped with greater consistency, thus enhancing robustness. Dimensions were normalized to generate a Composite Trust Index (CTI). This allowed us to be able to visualize that trade-off between the accuracy of long-term predictions versus how sustainable it could be. It is notable that the performance of PMI agents has a direct relationship with a large CTI: hence trust can be linked to the algorithm and greener process conditions. From the results, we know that Hybrid-UQ is the most reliable and sustainable among them with statistical significance.