Issue link: https://www.e-digitaleditions.com/i/1499609
30 Pharmaceutical Technology ® Trends in Manufacturing eBook May 2023 PharmTech.com Process control Industrial Internet of Things (IIoT)-enabled sen- sors collect data from hyperspectral imaging spec- troscopy (e.g., near infrared or Raman spectroscopy) and mechanical stress data (e.g., forces, vibration, sound, temperature) from haptic sensors. Gastone explains that these data are used to infer the state of the granulation process, including particle size and composition. "This [method] is an indirect method of measur- i ng pa r t icle size a nd d ist r ibut ion, as d i rect mea- surement in real-time is currently not feasible," says Gastone. "Our predictive controller takes the input from the IIoT sensors and uses machine learning a lgor it h ms, including deep rein forcement lea r n- ing, to adjust the process variables, such as impel- ler speed and liquid addition rate, in real-time to achieve the desired outcome. Ultimately, the goal is to control the particle size and distribution within the specified range." Developing the Luperini-Eigengran neural control system involved multiple machine learning tech- niques, from simple regressions to deep reinforce- ment learning, which involves training algorithms to make decisions based on experience. The AI algorithms in the cognitive pharma ma- chiner y were trained using a combination of his- torical process data and real-time data from the IIoT sensors, says Gastone. "The historical process data [are] used to build a physics-based model of the gran- ulation process," she further details. "This model is used to train the AI algorithms to predict the behav- ior of the process under different conditions. The AI algorithms are then trained using real-time data from the IIoT sensors to adjust and refine the model based on the actual behavior of the process. The AI algorithms are also trained using deep reinforcement learning, which involves rewarding the algorithms for making good decisions and penalizing them for making bad decisions. This [method] allows the al- gorithms to learn from experience and improve their decision-making over time." Data from the spectroscopy imaging and haptic sensors are crucial to this approach because they in- clude real-time chemical and physical information about the process. "During the development process, we used machine learning algorithms to analyze the data collected from the sensors and learn the correla- tions between the sensor readings and the endpoint of the granulation process. The incorporation of deep learning algorithms into our control system enabled the system to learn from experience and improve its predictions over time," explains Gastone. The system was able to autonomously detect end- points and provide real-time feedback to operators. Additionally, the control system can adapt to chang- ing conditions. This cognitive capability can provide more accurate and efficient monitoring and control of the granulation process, resulting in improved qual- ity and consistency of the final product. The proto- types of the cognitive pharma machinery control sys- tem in single-pot HSWG have demonstrated improved performance and accurac y compared to existing methods, reports Gastone. The researchers are currently working to extend the cognitive control system to encompass a HSWG unit operation going into a f luid-bed dr ying (FBD) process. Machine learning techniques are again used in these operations to analyze data and to train the system to recognize patterns and predict outcomes. The control system can learn to optimize granula- tion in the HSWG unit based on how the granules are subsequently dried in the FBD unit, explains Ga- stone. "This integration results in improved moisture control within the granules and optimized energy consumption through better thermal profiling," she says. The researchers are not currently looking at f luid bed granulation, which can be controlled using conventional control systems with PAT sensors. In HSWG-FBD, however, the AI-based system adds value because it can learn from multiple unit operations and optimize each individually to achieve the best results, says Gastone. Regulatory considerations Although AI is not specifically regulated, Gastone says that Eigengran's AI-based control system was developed to follow existing guidelines, such as the US Code of Federal Regulations (CFR) 21 Part 11 for elec- tronic records (3); European and US requirements for installation, operational, and performance; and the International Society for Pharmaceutical Engineer- ing GAMP 5 guidelines (4). "We have already initiated the validation process with our early adopters to ensure that our technol- og y is f ully qualified and validated before deploy- ment," says Gastone. "Our approach involves start- ing with our technology in process development to optimize the parameters, which can then be inte- grated into an approved control plan. We anticipate concerns from the regulatory side regarding the use of AI in the pharmaceutical industry and are confi- Cognitive capabilities provide more efficient monitoring of granulation.