Process optimization assistant powered with AI driven digital twin: multiobjective processes use case

Digital twin technology is used to create a numerical version of a physical process. To this purpose, we propose to use AI models, in particular deep learning, for their versatility and efficiency to extract patterns from data. Input parameters corresponding to process parameters are used to predict material characteristics, the outputs. Digitalization of a process brings several added values to engineers that can explore and better understand their process at a greatly reduced cost. Our module of optimization gives process engineers the possibility to quickly find the best configurations of a process depending on their specifications. With an AI driven digital twin, physical simulation can be used in complement of data science models to extend the data and validate the models. Most processes require several outputs to be optimized. The straightforward approach is usually to combine outputs within a formula, named a figure of merit, hiding the complexity of the process. This approach is difficult to tune as models must be retrained at any output weighting modification. We propose to use multi-objective optimization to have an ensemble of optimal solutions. Optimality is defined by the Pareto front property ensuring that in any pair of solutions, a solution is performing better for one objective and the other solution for the second objective. This property makes the process engineer essential in the choice of the solution to appropriately decide which tradeoff to apply between objectives. Additional constraints and modifications in output importance can be added at optimization stage. Finally, this pipeline can be used by process engineers as guidelines to explore new experiments based on relationships discovered with the AI models. More complex DOE (Design Of Experiments) can be explored easily within AI models, as complex interactions are better explained, helping process engineers to gain stronger knowledge on their process with fewer experiments compared to traditional methodology.