A collaboration between scientists at Imperial College London and Swiss data analytics developer DataHow has created a new software function now available for Design of Experiment (DoE) for upstream processes.
The new sampling algorithm can not only use existing knowledge to propose experiments that can improve a process, but it can also provide this information in an explainable way.
“The key message I want people to take away is that doing DoE based on a trained model is more efficient than classical DoE methods, or tools that use one factor at a time,” says Sam Stricker, a PhD student at Imperial College, who developed the new model.
According to Stricker, unlike classical Bayesian optimization techniques, which made a simple trade between a couple of variables, the pareto-front-driven algorithm can find every trade-off between multiple factors.
“In classic Bayesian optimization, if you want to minimize aggregates and maximize titer, you take the upper confidence bound for titer and the lowest for aggregates and then decide if you might be willing to accept slightly higher aggregates for a higher titer,” he explains.
“But with pareto front, you take every point for each titer level, and then each aggregate level, and then make a trade-off about how to weight them. That means you can, for example, reject experiments where any improvement in titer is going to be smaller than the measurement noise.
According to Stricker, the new algorithm is already incorporated in DataHow software and can currently be used for upstream processing. A proof-of-concept of using the algorithm downstream, in chromatography, is currently in the test phase.
“From a DataHow point-of-view, they’re working on constant optimization and they’re very focused on the practical side of things—making it useful for industry,” says Stricker.
Stricker is dedicating the rest of his doctoral research to improving the algorithm, including making it more accessible to non-data scientists.
“DataHow wants to democratize machine learning tools, so everyone can model processes to get the insight they need, so I’m working on making the algorithm more explainable and easier to understand,” he says.