Problem
A large chemical enterprise was looking to optimise the configuration of the production process used in the manufacture of a new product. The aim was to maximise the quantity of the product produced, while minimising the cost of the production, and preserving form properties of the product. Due to the complexity of the process, it was challenging to predict the result of combinations of parameters in the production. Each trial of a new configuation for the production process was also costly and time consuming, and therefore a systematic approach to this optimization was required. The product was costly and time consuming to manufacture meaning that the iteration count needed to be minimised as much as possible.
Solution
There were many examples of previously run configurations in similar production processes to that which the customer was interested in. These previous experiments highlighted significant nonlinearities in the combined behaviour of the complex system. This made it highly challenging to infer the next best configuration to trial in order to balance the exploitation of promising areas, but also the exploration of large areas of uncertainty towards the extremities of each parameter.
Working with the customer using our in house Bayesian Optimiser, we were able to effectively encode the configuration space and constraints of the problem, and define a scoring function encapsulating all quantities of interest. Seeding the optimiser with existing information from similar production processes quickly led to much improved results in the process, in far fewer iterations than would have otherwise been required, saving a significant amount of time and money for the client, and delivering a result that was better than what would otherwise have been achievable.
