Session: Government Agency Student Posters

Paper Number: 173041

Optimizing Machine Learning Models for Real-Time Ultrasonic Quality Monitoring in Pharmaceutical Tablet Manufacturing 

The pharmaceutical industry is increasingly embracing advanced quality control strategies to support Real-Time Release Testing (RTRT) and Continuous Manufacturing (CM). Yet, as material systems and quality expectations grow more intricate, the development of accurate, model-based evaluation frameworks for Oral Solid Dosage (OSD) forms has become a critical challenge—frequently creating a bottleneck in the quality assurance process. Ultrasonic wave propagation offers a promising solution to this issue by enabling the extraction of viscoelastic and scattering properties of OSD forms nondestructively, quickly, and cost-effectively. These properties can inform the user about the chemical/mechanical makeup and the existence of any internal defects in the OSD. However, for most practical applications, this process proves to be problematic due to the extremely difficult nature of the inverse wave propagation mathematical problem that arises from trying to extract the Critical Quality Attributes (CQAs) of the pharmaceutical tablets. Specifically, recent research has focused on extracting material attributes such as the stress relaxation time, strain relaxation time, and scattering coefficient in the Rayleigh regime. To use ultrasonic evaluation applying first principles, many assumptions have to be made that can influence the accuracy of the results, especially as the compositions of the OSD forms grow increasingly complex. Recent advances have introduced PQCNet models, which are Multi-Output Regression (MOR) machine learning (ML) models, that extract the CQAs from the raw ultrasonic waveforms. Building on recent advancements in ultrasonic and ML approaches for characterizing compressed Oral Solid Dosage (OSD) forms, this work aims to build on the ML-based material property prediction through integrating the Optuna framework to optimize the ML models’ hyperparameters. The resulting models will balance the pharmaceutical industry’s need for accuracy with performance limitations, which will be measured in floating point operations (FLOPs). Some examples of hyperparameters in PQCNet models being optimized by Optuna throughout this study include the number of layers, nodes per layer, number of epochs, batch size, and learning rate. Waveform data to train, validate, and test these models is collected under varying compaction conditions using advanced laboratory set-ups from the Photo-Acoustics Research Lab at Clarkson University. The data collected experimentally will also be supplemented by synthetic datasets generated with Wolfram Mathematica to enhance PQCNet model robustness. The ultimate goal is that the study will help develop a scalable, realistic, automated pipeline for RTRT, CM, and in-line quality monitoring of CQAs in pharmaceutical manufacturing. Preliminary results have found the optimal number of hidden layers, batch size, and learning rate for the PQCNet models based on the current datasets. This allows us to focus in on the other hyperparameters in the future as we continue to fine-tune our PQCNet models. The results found when attempting to minimize error improved the results over existing ML models, lowering the average error from 1.6% to 0.33%, which is nearly a 80% drop. Furthermore, the results have shown that the previous assumption that a funneling ML architecture was optimal is false. Using this process, a maximum FLOP count can be input into the code based on the limitations of the given hardware, and the software will generate the most error-free model possible within the limitation. This will result in a framework that can be applied across pharmaceutical quality control with a wide variation OSDs. The resulting ML models can then be implemented efficiently and cost-effectively into the quality control process and aid in the FDA’s goal of transitioning to CM and RTRT.

Presenting Author: Christian Di Spigna Clarkson University

Presenting Author Biography: Christian Di Spigna is an electrical engineering student with a strong background in mechanical engineering at Cedarville University, where he is a President’s Society Fellow. He has completed both an industry internship at New Scale Technologies as well as a research internship at an NSF REU at Clarkson University. During his internship at New Scale Technologies, Christian gained hands-on experience which included manufacturing motor cores, assembly testing, sensor technology, and PCB design. At Clarkson, he worked with mechanical engineering problems involving machine learning, continuous manufacturing, and ultrasonic data acquisition on pharmaceutical tablets for the purposes of quality control. Christian also received the opportunity to present his research at the summer RAPS event, along with an invitation to the prestigious TECHCON 2025. Christian has extensive experience with MATLAB, Simulink, Altium, SolidWorks, Python, C/C++, and VHDL. He dedicates his time to several engineering competitions, having been a member of the ASEE robotics competition team as an underclassman before shifting into a mentor role. Additionally, Christian is currently working on his senior design team as a member of the ECE design team on the NASA Student Launch competition. He currently serves as the treasurer of the Ohio Nu chapter of Tau Beta Pi and is responsible for running their weekly tutoring nights for underclassmen engineering students. Christian also is an active member of Cedarville’s undergraduate interdisciplinary honors society, Tau Delta Kappa. Upon graduation, he plans to take the FE exam and pursue graduate studies in either sensor technology or robotics. In the long-term, Christian plans to advance cutting-edge technological innovations in the sensor and robotics fields.

Authors:

Christian Di Spigna Clarkson University
Tipu Sultan Clarkson University
Cetin Cetinkaya Clarkson University