Forecasting Algae Growth in Photo-Bioreactors using Attention LSTMs

Sustainability is the current global challenge. This is reflected not only in the demand for healthy food but also in the objectives of CO2 neutrality as an influencing factor for producing companies. These challenges can be met with the industrial cultivation of algae: Algae have a high potential to be used not only as food supplements, nutraceuticals or in pharmaceuticals but also as biological fuel. In particular, their ability to bind CO2 during growth also makes them suitable for use as CO2 sinks. This ability is enhanced in particular by the fact that their relative yield density per area is higher than that of other plant species. Despite the high potential of algae, there are currently limitations to their large-scale use, as scaling up from laboratory environments to pilot applications typically requires more than 5 years. This challenge is characterized by the fact that the growth behavior of algae is mapped by highly complex interactions: As biological systems, algae are not only influenced by current environmental conditions but also by past environmental conditions. These interactions make current pilot applications inefficient due to insufficient control and monitoring techniques. However, this limitation can be countered by developments in the context of Industry 4.0: By using modern communication and evaluation technologies, a “smart” bioreactor can be developed, which evaluates algae growth in real-time, performs process adaptations and thus significantly accelerates algae growth and scale-up. Because of this background, an algae bioreactor was established at the Centre for Advanced Manufacturing at the University of Technology Sydney, which enables the analysis of algae growth, in particular through machine learning. The subject of this paper is the study of algae growth using LSTMs. In order, to learn the biological behavior of algae in the shortest possible experimental environments, a series of experiments with repetitive change intervals were run by systematically varying the environmental parameters. However, since algae growth exhibits high historical dependence, LSTMs were trained to model algae growth. These types of recurrent neural networks profit from a memory cell architecture that allows using information dating back very far in the past, due to the optimized gradient flow.
To model the importance of influencing factors attention mechanism is used on both, the variable and the temporal direction. The LSTM is compared to a transformer model and a statistical ARIMA model. Based on the trained models, the behavior of algae growth in the domain context is interpreted: This shows that in particular, bubbling CO2 into the algae has a high effect on algae growth during the series of experiments. However, the trained models show that CO2 uptake by the algae is not increased. This leads to the hypothesis that adding CO2 results in intermixing of the algae biomasses, thus realizing a more uniform light distribution within the biomass and positively affecting the growth rate. In further tests, this hypothesis will be further tested and an algorithm will be derived from the trained models, which independently adjusts the parameters of the bioreactor to positively influence the algae growth.