What Can We Do to Advance the Biomanufacturing Process?

    Manufacturing within cells is extremely complicated and can easily go awry while on Earth, let alone in space.

    One of the key issues that we need to take into consideration is manpower, which is extremely valuable in space. Yet throughout biomanufacturing design and operation, the space industry sector does not have analytical tools capable of monitoring how cells are coping with the stresses and strains of production. A lot of factors including temperature and culturing conditions affect bacterial growth, and without these cell analytics we cannot react to in-process issues in time to keep manufacturing on track.

    Therefore, a tool that can take multiple factors into consideration with high efficiency and the development orientation of unmanned environments is essential for us.

    We came up with the idea that artificial intelligence can fulfill our needs. Our ultimate goal is to let the system pick out the colony of bacteria most suitable for our causes under any conditions that we want without an operator beside it when it works.

Why Artificial Intelligence?

    The reason why our software used AI technology is that, it is the best image processing method in existence; and on the other hand, the more it is trained, the "smarter" it will become. It is bound to surpass human beings and the existing appliances’ identifying method one day in the near future.

    Conducting experiments in space are bound to be automated. One of the most important parts is computer vision, while the most mature technology is deep learning.

    AI, coupled with Internet of Things [IoT] sensors and Digital Twins (process models), is driving a fourth Industrial Revolution offering higher yields of better quality products from more efficient processes. A fast, rich cell analytic is vital for biomanufacturing to join the fourth Industrial Revolution. We, NCKU_Tainan are developing a technology to drive next-generation biomanufacturing.

Deep Learning vs. Machine Learning

    Deep learning is a subset of machine learning. They are both implementations of artificial intelligence, imitating decisions of human beings.

    The goal of machine learning is to find an appropriate function capable of dealing with corresponding inputs and outputs. In most situations, machine learning solves problems with correct answers, such as classification. Humans feed machine learning models with structured data as input, in other words, we properly identify the input features. For example, when tackling animal classification issues, the machine might receive a set of inputs, ‘it is flurry,’ ‘sharp teeth,’ ‘retractable claws,’ ‘long tail,’ ‘four legs’ and some other information, and the output could probably be a cat. These are structured data with specific information.

    On the other hand, deep learning is an evolution of machine learning, extracting features on its own, learning the relationship, and learning from its own errors. Structured data is not necessary, so in the animal classification case, deep learning might just receive an image of a cat. It observes on its own about which part of the image is meaningful or learns about the relationship between specific pixels. This evolution is due to its architecture. It uses a layered structure of algorithms called Artificial Neural Network (ANN) to mimic the way a human would think in daily life. In much simpler terms, it replicates the operation of a human brain, and passes data through processing layers to interpret data features and relations.

Table 1. Example of animal classification

    In conclusion, machine learning requires structured data, which means more human intervention, and deep learning does not. Therefore, deep learning is capable of dealing with higher dimensional input from more complicated issues, requiring much more calculations, time consumption and better resources. However, deep learning usually requires more computing power (GPU) in the training process, while machine learning can be done with CPU.

What Have We Done So Far?

    We speculate that the survival rate should be related to the concentration of selenomelanin, the darker the color, the higher possibility for the colony to survive under radiation. To verify the function of our Se coli with its survival rate, the first step at this moment is to apply deep learning on analyzing image data regarding colony color and size. Therefore, we decided to input the experimental photos and their corresponding data to model training.

    After previous experiments that apply deep learning to E.coli analysis, we observed that despite the small database, by improving our data set, retraining our model, and repeating, the model would be accurate enough. We can use the results for deployment design combined with a user interface (UI) to achieve a fool-proof operation mode (Input photo of colonies→ predict its utilization value)while other environmental factors are fixed.

    By picking out the colony with the darkest color and most suitable radius with artificial intelligence, we realize which colonies are more likely to successfully manufacture the applications that we want in space .

    Furthermore, we would like to highlight that considering multiple factors at the same time and picking out the best strain is not the only advantage. By getting more and more data on every aspect that would affect the growth of our Se coli or other types of microorganisms in the near future, our system can learn from these previous experiences and expand its database , which means that our prediction accuracy will keep on improving forever.

    In the future, we can use our current software as the preliminary design, and then go further to develop other possible advanced tasks. (eg: Using AI tools to automatically pick out the best strains.)

Fig. 1. Testing out our model using Landing AI’s system with live deployment

Challenges of Introducing AI in Biomanufacturing

    The following content is the difficulties we may experience with delivering efficient end-to-end AI-based solutions in biomanufacturing.

Data Quality

    Since the complex environments generate high data volumes that need to be preprocessed and analyzed by Data Scientists to create meaningful results, poor data quality is a significant obstacle in data-driven biomanufacturing.

Data Quantity

    Although biomanufacturing processes produce lots of data, it is common that the insights laid in data is either not adequately collected or the process lacks a suitable infrastructure to handle large-scale datasets.

Infrastructure Challenges

     The industry requires new tools and strategies to support the complicated process. Software that supports the operations is distributed and does not cover processes as a whole. Connectivity and the ability to generate a smooth manufacturing process is what urgently needed to be solved.Connectivity and the ability to generate a smooth manufacturing process is what urgently needed to be solved.