Materials design with MLatom for ammonia separation and storage

Materials design with MLatom for ammonia separation and storage

We used our MLatom package to develop a machine learning approach for designing materials based on mixed metal halides to facilitate ammonia separation and storage.

Ammonia is one of the most important industrial products as it is crucial for synthetic fertilizers which help to feed humankind. Ammonia is produced via a very energy-intensive, large-scale Haber−Bosch process which traditionally relied on a stable electricity supply. The development of intermittent renewables creates a major challenge for this process which requires solutions adapted for smaller-scale production. One of the suggested solutions is to improve the efficiency of the process by separating (and storing) ammonia from the reaction mixture with metal halides and later release ammonia. This requires finding conditions for a right balance between sorption/desorption energies and choosing appropriate materials is crucial. A promossing type of materials are mixed metal halides, but a number of possible combinations of metals and halides is combinatorially very large as we can choose different fractions of both methal and halide in mixed compounds. Both experimental screening of materials and calculation of desorption energies with traditional quantum chemical methods would require lots of resources and become infeasible for large search spaces.

In our collaborative work, we explored mixtures of Mg, Ca, Cl, and Br, which can absorb different amount of ammonia. The number of mixed metal halides even for such a limited selection of components was 4096 and each combination can absorb 1, 2, or 6 ammonia molecules. Our collaborators first established a robust computational scheme with density functional theory calculations to calculate the desorption energies. Then lots of effort was spend to building a machine learning model which can reliably predict desorption energies. The major step was finding appropriate descriptors and after screening and filtering many of the structural and quantum chemical descriptors we identified the most important ones. One of the most important descriptors (polarizability) requires however quantum chemical calculations which would slow down the computational screening. Thus, we also machine learned this descriptor based only on structural information.

In the end, we saw that only 45 training points were remarkably sufficient to build a reliable machine learning model predicting desorption energies with chemical accuracy. We expect it to be applicable to a wider range of metal halides.

All machine learning calculations were performed using our MLatom package which can be used without any installation via a web browser using our MLatom@XACS cloud computing service. We also provide user support.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.