NEU-2003

The new technologies in chemical industry require quick response on market necessities. It needs deep investigation of properties of new materials for a short term. Moreover, it requires a lot of experimenting. The problem of reduction burden and cost of experiments is quite urgent.

We developed specific software tool for the decision of such problems. It is a code for working with an experimental database and predicting material properties subject to used ingredients.

The code is applied to create non-linear regression model of existing empirical data by using neural networks. Adapting neural network model to the data, one balances between Scylla of over-simplified model with poor approximation and Charybdis of over-complicated model adapted to data noise and thus with poor generalization capability for previously unseen data. The developed software successfully decides this problem. Comparing with other techniques, neural nets are especially effective in the case of many dimensions (i.e. many ingredients).

There are some distinctive properties of this software:

• The code controls the smoothness of the response surface by using committee technique and Laplace regularization. These methods improve generalization of model predictions.
• Committee technique also supports the stability and repetition of model constructing.
• Laplace regularization is used to avoid model over-fitting and supplies the robustness.
• The model also estimates noise by reporting the confidence of its predictions.

All these features support the opportunity of code usage for an inexperienced in neural networks user.

All experimental data, accumulated by the company so far, are involved in the development of new products. Neural networks construct informational model using experimental data. This model is used to create artificial data and predict some properties of these data. The result is considerable cost savings and decreasing of new products development cycle.

The constructing model can be used for solving a whole series of problems.

• User can obtain not only the simple result for one point but investigate its vicinity and understand the stability of examined properties.
• The model contains the measure of its accuracy, so user can determine the points for new experiments and their necessity.
• It can be used for solving inverse problem – the creation of material with specified properties.

The developed code is a useful tool for chemical laboratory. It can be applied by an inexperienced in neural networks specialist. It is effective in problems with a big amount of experimental data and can reduce number of experiments. It helps to investigate various mixtures in pharmaceutics, food, health care, and beauty care industries depend on a proper formulation of their ingredients.