2025 Spring Meeting and 21st Global Congress on Process Safety
(167b) Adaptive Experimental Design for Fermentation Process Optimization
Valuable natural products are often produced through complex bacterial fermentations that require media with multiple ingredients and complex sets of control conditions. These multiple inputs make optimization of the fermentation process challenging for classical design of experiments, which requires many condition combinations to characterize input interactions. We have implemented machine learning algorithms, including Bayesian optimization, in an adaptive experimentation method to generate novel fermentation process recipes that efficiently drive experimentation towards an optimum recipe. Our initial efforts in medium-scale lab testing quickly identified highly productive recipes that were unsuitable for large-scale production due to accompanying flaws. We addressed these flaws by balancing productivity goals against ingredient costs, viscosity, and risk of cell lysis. This method eventually yielded a recipe that achieved productivity improvement averaging more than 7% over the current standard in repeated manufacturing batches at equivalent input costs. Additional recipes suggested by further rounds of adaptive experimentation may result in greater improvement, since limited resources have been invested so far. These results demonstrate that adaptive experimentation can accelerate natural product fermentation optimization over traditional methods.