A multidisciplinary research team used machine learning and physics to characterize the rough particle surfaces in powders and pills used in pharmaceuticals. 

"AI in medicine promises to provide composite, panoramic views of individuals' medical data; to improve decision making; to avoid errors such as misdiagnosis and unnecessary procedures; to help order and interpret appropriate tests; and to recommend treatment." - Eric Topol.

An engineering and research team is creating a revolutionary AI-based estimate for producing medicine utilizing physics and machine learning. The goal is to boost productivity and accuracy so fewer product batches fail.

Overview

Machine Learning can now image scattering mediums and suppress speckles. In both instances, the speckle pattern is regarded as an undesirable disturbance. The extraction of spreading media information from the speckle has also been pursued qualitatively, such as classifying materials based on their dispersed speckle patterns. The main difficulty of quantitative speckle analysis is the phase signal's sensitivity to surface randomization, which inhibits neural networks' capacity to discern other dynamics.

Challenges

When pharmaceutical companies make pills and tablets to treat various ailments, aches, and pains, they must separate and dry the active medicinal ingredient from a suspension. A human operator is needed to monitor an industrial dryer, stir the material, and wait for the compound to take on the proper properties for compressing into medicine. The operator's observations are crucial to the work.   

Physics and machine learning are used to classify mixed particles' rough surfaces. The technique, which employs a physics-enhanced autocorrelation-based estimator (PEACE), can revolutionize pharmaceutical manufacturing procedures for pills and powders, enhancing efficiency and accuracy while resulting in fewer failed batches of pharmaceutical products.  

Approach

To determine whether a compound is adequately mixed and desiccated in the pharmaceutical industry, it is customarily necessary to halt an industrial-sized dryer and remove samples for testing. Researchers at Takeda believed that artificial intelligence could enhance the task and reduce production-slowing stoppages. Initially, the research team intended to train a computer model to supplant a human operator using videos. However, the determination of which videos to use to train the model remained subjective. 

Machine learning characterizes particle sizes, and physics specifies the laser-mixture interaction. Physics provides rapid neural network training. Thus the machine learning algorithm only needs a few datasets. Experts can ensure their experiments' real-world consequences by merging both parties' knowledge and goals. The team plans to file a third patent.  

Conclusion

Getting numeric information from an imaging system about surfaces that scatter light a lot is complicated because the phase of the scattered light changes many times as it travels, creating complex speckle patterns. One specific use is drying wet powders in the pharmaceutical business, where measuring the particle size distribution (PSD) is essential. There needs to be a non-invasive, real-time tracking probe for the drying process, but none works. 

In this report, the researchers create a theoretical relationship between the PSD and the speckle image and describe a physics-enhanced autocorrelation-based estimator (PEACE) machine-learning algorithm for speckle analysis to measure the PSD of a powder surface. This method answers both the forward and backward problems at the same time. Furthermore, the physical law governs the machine learning approximator, making it more straightforward.

Image source: Unsplash