A new artificial intelligence system has been developed by the team of researchers that can pore through scientific papers as well as extract recipes in order to produce particular types of materials.
This work has been envisioned as a first step towards the system by the team of researchers which has the ability to originate recipes for those materials which have been described only theoretically. A further step has been taken in that directions by the team of researchers along with a new artificial intelligence system which can recognize the higher level patterns that are consistent across recipes.
This new system has the ability to identify correlations between precursor chemicals which have been used in material recipes as well as the crystal structures of the resulting products. The same correlations have been documented in the literature.
Also statistical methods have been considered by this system through which natural mechanism has been provided in order to generate original recipes. This mechanism has been used by the researchers in order to suggest alternative recipes for known material and the suggestions have been accorded well with real recipes in the paper.
This new system of team of researchers is a so called neural network like most of the best performing other artificial intelligence system of few past years. By analyzing broad sets of training data, it learns to perform computational tasks. Traditionally in order to generate material recipes, the attempts have been used by neural network have run up against two problems which have been described as sparsity and scarcity by the researchers.
For a material, any recipe is able to be represented as a vector which is important for a long string of numbers. A feature of recipe has been represented by each number for example: concentration of a particular chemical in which the solvent has been dissolved or the temperature as well at which a reaction takes place.
There will be most of the numbers zero if only a few of the many chemicals and solvents which have been described in the literature will be used by any given recipe. This has been described as sparse by the researchers.
Similarly in order to learn modification reaction of parameters for example: chemical concentrations and temperatures, final products can be affected which is a system that ideally be trained on a big number of examples in which those parameters have been varied. A few recipes can be contained by the literature only for some materials especially newer ones and that has been termed as scarcity by the researchers.