Connectionist Models in Materials Science: Characterisation of the Sorption Properties of Hydrogen Storage Materials

We explain how connectionist models can be used in research and development in the areas of materials to make it more productive and useful at much lower cost and time. The basic idea is to identify a computational model using neural networks to characterise the relation between the output characteristics, input ingredients and process parameters. As an illustration, we focus on the problem of characterising the sorption properties of hydrogen storage materials. We consider the composite materials La ₂ Mg ₁₇ - x wt% Z with Z = LaNi ₅ and Z = MmNi _4.5 Al _0.5 for various values of x. We use training data on the desorbed amount of hydrogen for two different temperatures and different time of desorption. These training data are used to train a multilayer network which is then used to predict the amount of released hydrogen for new desorption temperature, and desorption time. Our results show that for both materials with different values of x, the network is able to learn the non-linear desorption characteristics quite successfully. Hence, for different temperatures and desorption time, we are able to predict the dehydriding kinetics and storage capacity of hydrogen storage materials without doing the actual experiments.