Purpose
Current bioplastic production using polyhydroxyalkanoates (PHA) is limited by production costs associated with maintaining a pure culture and using synthetic chemical-based feeds. Significant barriers for commercialization of bioplastics associated with high production costs are maintaining sterile conditions for a pure culture and the acquiring/synthesizing chemical-based feeds. This research uses a more practical approach that substitutes synthetic feeds for a more robust manure derived Volatile Fatty Acid (VFA) rich feed. VFAs are acquired from acidic fermentation of organic matter (dairy manure) and are important precursors to PHA synthesis by bacteria. Another change is the use of a mixed microbial culture (MMC) instead of a pure culture. MMCs are a collection of many types of bacteria compared to pure cultures which aim to only contain a single species. Usage of VFAs and MMCs help reduce the main drivers behind high production costs by alleviating the need for sterile conditions and expensive feeds. However, for a commercialized process there is another significant barrier that needs to be overcome. Real-time monitoring of VFAs remains elusive while being essential to commercialization. Without the ability to monitor VFAs in real-time results in the inability to confidently maximize PHA yield on VFA-based feeds.
This research establishes a more efficient and reliable alternative to measure the concentration of VFAs in MMCs. Ensuring VFA depletion is essential to maximize PHA yield. The current method requires a methodical examination of VFA uptake by the consortium through measuring volatile suspended solids (VSS), VFA concentrations, and duration of examination (time). This is a time intensive process and is not always accurate. Without real time, measurement of VFA concentration the consortium can experience a buildup of VFAs instead of depletion as intended leading to a decrease in PHA yield. This research uses machine learning to predict the concentration of VFAs in real time, alleviates the time requirement needed, and has potential to minimize the chance of buildup instead of depletion. Additionally, using a real-time measurement strategy is more akin to a commercialized process, the end goal for PHA production.
What Did We Do?
The first step towards building a reliable machine learning model was to collect data that could be used to train a model. For this we used ten experimental production runs (an experiment that feeds VFA rich substrate to a mixed microbial consortium to build up PHA concentrations internally) measuring run duration, dissolved oxygen (DO), Oxidative Reduction Potential (ORP), Oxygen Uptake Rate (OUR), and Waste Activated Sludge (WAS) Concentration. DO is the measure of oxygen dissolved in the water and OUR is the measure of the rate at which oxygen is being consumed by the consortium. ORP is the measure of how likely a medium is to give or receive electrons in a redox reaction in wastewater in can be a measure of the reactors condition (aerobic, anoxic, or anaerobic). Lastly, WAS Concentration is the measure of percent WAS used to start an experiment based on a maximum value of one liter. Using this data we also calculated some additional measurements before the data reached the model to give more insight into reactor’s condition throughout time however, these were less impactful than the variables above. This data was split into a training set (Ten out of 16 runs) and a testing set (Six out of 16 runs). With this data we utilized five main machine learning models (linear regression, support vector regression, decision tree, random forest, and neural network) to make accurate and precise predictions of VFA concentration.
We differentiated the different models by using three primary statistics: mean squared error (MSE), mean absolute error (MAE), and r2. MSE is the average squared error between the models’ predictions and the measured concentration. MSE was chosen as it more heavily penalizes larger errors due to the squared nature of the statistic. MAE is the average error between the predicted and measured concentrations. MAE was chosen as it gives a more interpretable measure of the models’ accuracy. This metric is not influenced by large outliers and treats over and under predicted errors similarly and overall is easier to interpret compared to MSE. The r2 metric is a representation of the model’s predictive accuracy. This metric was used as it gives a numerical representation of a model’s fit to the measured concentrations The models were evaluated using a single measurement, combination of measurements, and lastly all four measurements (Time, DO, ORP, and OUR). The optimal model was determined by the combination of model and parameters that resulted in the lowest deviation from the measured concentration and highest r2.
What Have We Learned?
This research has allowed us to increase production run optimization by more accurately predicating VFA concentrations in real time. Real time prediction is the beginning step to the realization of a reliable commercialized process. Bioplastic production using PHAs is primarily limited by cost of production. Using a resource such as dairy manure to produce VFA rich substrate can minimize the production cost but without real time monitoring can lead to excess costs. The biggest benefit of using machine learning to predict VFA concentrations is its ability to minimize PHA production costs by accurately predicting VFA depletion. The model being used currently is a random forest model using all parameters (DO, ORP, OUR, WAS Concentration, etc.). Using this model on test data results in an r2 around 0.89 and MAE of 1.6 Cmmol/L with some variation depending on the testing set used.
Future Plans
Using a machine learning model to monitor real time measurements for predicting VFA concentration is to be utilized in operating an autosampler. The model would be used to automatically take samples from and feed the production reactor with VFA rich liquor. This setup will be assembled to mirror a commercialized process in which PHA is produced autonomously with operators overseeing the process. Conducting automated production runs will allow us to investigate the upper limits of a mixed microbial consortium to produce and store PHA. This will give us insights into the optimized amount of VFA liquor to produce for a predetermined reactor volume.
Authors
Presenting & corresponding author
Brandon M. Boyd, Research Assistant, University of Idaho, Boyd4708@vandals.uidaho.edu
Additional author
Dr. Erik Coats, Associate Professor, University of Idaho
Acknowledgements
This research was funded by the USDA Sustainable Agricultural Systems Initiative through the Idaho Sustainable Agriculture Initiative for Dairy (ISAID) grant (Award No. 2020-69012-31871).
I would like to acknowledge my coworkers who have helped me conduct my research, Dr. Moberly in the chemical engineering department and JR in IT for helping me with my Machine Learning model and future plans, and lastly Dr. McDonald.
The authors are solely responsible for the content of these proceedings. The technical information does not necessarily reflect the official position of the sponsoring agencies or institutions represented by planning committee members, and inclusion and distribution herein does not constitute an endorsement of views expressed by the same. Printed materials included herein are not refereed publications. Citations should appear as follows. EXAMPLE: Authors. 2025. Title of presentation. Waste to Worth. Boise, ID. April 7–11, 2025. URL of this page. Accessed on: today’s date.
