Bioprocess Optimization Based On Historical Data and Artificial Neural Networks
For all established biological processes, a wealth of data exist, whether in written form, electronic media, or in anecdotal form via process operators. Many times this data remains an under-utilized resource because efficient methods of capturing the knowledge represented by this data do not exist. By developing computational methods to examine archival data, it should be possible to use the natural and planned variations in past processing to optimize future fermentations. The methods developed in our lab will provide a way of identifying and repeating beneficial behavior while avoiding conditions that favor poor productivity or quality, thereby fully utilizing the knowledge-base for the process. Traditional experimental optimization methods such as factorial design experimentation and response surface methodology at production-scale are limited by high cost and equipment availability, as well as regulatory issues in the case of biopharmaceuticals. Thus far, we have used wine processing as a model system and have generated historical data on approximately 250 lots of wine (Sauvignon blanc and Cabernet Sauvignon) over the past three vintages. We have used this data to establish neural network training methods for this type of process data and to demonstrate that we can use the trained neural networks to model fermentation kinetics, as well as chemical and sensory attributes of the finished wine. We have also adapted both gradient-type and stochastic (e.g. simulated annealing and genetic algorithm) optimization methods for use with trained neural networks. Our current research is focused on finding methods for searching large databases of information for the most critical processing inputs, as well as extending the methods established in our lab to time-dependent data and to fermentations producing recombinant protein.
Technology for the Prediction and Prevention of Stuck and Sluggish Wine Fermentations
Stuck and sluggish alcoholic fermentations are an important problem in the wine industry, as the residual sugar left in the wines from these fermentations poses a potential stability problem in the final product. The ultimate goal of this project is the development of methodology for prediction of the kinetics of both normal and problem wine fermentations, based upon juice characteristics and intended processing. There are two components to this project, generating data on the impact of various parameters on fermentation kinetics from defined studies that will lead to a mechanistic model of cell growth and sugar utilization and use of that information for the training of neural networks that can predict fermentation behavior. This work is part of a multi-disciplinary effort in the department along with Drs. Bisson, Butzke, and Mills.
Biological Control of Plant Diseases and Methods for Efficient Process Development
Fungal diseases of plants pose a serious economic problem for agriculture in California and throughout the country. Eutypa Dieback in grapevines is one of these diseases that is typically controlled by repeated application of chemical pesticides, or in extreme cases, by removal and replanting. We are collaborating with Prof. Jean VanderGheynst in Biological and Agricultural Engineering to find new methods for the production of biological control agents that may increase their efficacy in the field and novel means for applying them in the field. We are currently working with Fusarium lateritium which has been shown to be active against Eutypa lata, the causative microorganism for Eutypa Dieback. In our lab, we are focusing on finding efficient experimental optimization methods for simultaneously finding the optimal combination of media components (type and concentration), inoculum characteristics (such as strain, age, and size), and fermentation parameters (e.g. DO, pH, temperature, agitation rate, and aeration rate). To accomplish this, we are using a combination of statistical and artificial intelligence techniques.