19 October 2023
Based on our technology from design of experiments and machine learning, divis intelligent solutions GmbH develops tools for optimising Xylan enzyme cocktail formulations. They take availability and cost of the enzymes into account, which are both top priorities for producing innovative Xylans. Based on experimental results, predictive models are trained for optimising enzyme formulations for important Xylan applications.
EnXylaScope partner divis intelligent solutions GmbH is specialised in consulting, development, and deployment of cutting-edge Artificial Intelligence solutions for our clients in industries such as consumer goods, automotive and chemistry. Based on our software modules for machine learning, predictive analytics and optimization, we develop customised solutions for our clients. We typically work with our clients on product development in their R&D department and with the production department for data-driven process optimisation, predictive maintenance, and predictive quality. divis has a team of highly skilled and experienced experts in predictive analytics, statistics, machine learning, optimisation, and software development.
One of divis’ technology application domains in industry is in product development when the product is a formulation of ingredients, aiming to achieve certain desired properties. Examples include the formulation of cosmetics products for maximising shelf stability of the resulting emulsion, for achieving a desired microbiology performance, or for obtaining the target sun protection factor. Oral care, decorative cosmetics, and sensory profiles in food development are other representative domains in which our customers are active. Two sample applications are explained in more detail in the corresponding white papers with Beiersdorf and Johnson & Johnson.
Our approach is based on experimental data about formulations that have already been made by our customers. If no such data is available, we develop an experimental plan for our customers. Such a plan defines a minimum number of experiments to be conducted such that the information gain is maximised and a predictive model can be trained on the resulting data. Such a predictive model is then created by divis, using our advanced machine learning technology. The resulting predictive model can be used to predict, for a new formulation, the resulting properties (e.g., its stability). It can also be used to optimise a formulation towards best possible performance concerning a given property (e.g., longest stability).
For EnXylaScope, we work with the project partners to develop experimental plans for enzyme formulations that facilitate Xylan debranching processes. A realistic experimental plan is developed per feedstock type (e.g., poplar chips and wheat straw) that also takes availability and cost of the enzymes into account. This experimental plan is then executed by the EnXylaScope project partner Celignis, measuring the corresponding Xylan debranching properties for each formulation. The enzyme formulations and their debranching properties are then used as input data for our machine learning technology. Machine learning will yield a predictive model of Xylan debranching which, for a new formulation, will enable users of the model to predict its Xylen debranching capabilities. Combined with an optimisation algorithm, the model will also facilitate finding an optimal enzyme formulation that yields a given, desired Xylan debranching result.
The goal of our contribution to the project is to embed the above technology within a software toolbox that enables end-users to predict the right enzyme formulation for their desired Xylan-based product. Such a toolbox will be developed based on our proprietary ClearVu Analytics software for predictive analytics and optimisation.