Utilization of Artificial Neural Network to explore the compositional space of hollandite-structured materials for radionuclide Cs incorporation

Abstract Hollandite with the general formula A2B8O16 is known for its potential to immobilize radionuclide Cs in the tunnel along the z-axis of the crystal structure. The effective Cs incorporation in a hollandite phase with an optimal loading capacity and the long term stability depends significantly on the B-site cations, which, in addition to providing optimal structural compatibility, must ensure the phase's resistance to chemical weathering in an aqueous environment that includes external thermodynamic conditions such as temperature and solution chemistry. Based on the importance of the B-site cations, we explored in detail the possible B-site compositions by employing Artificial Neural Network (ANN) simulations and crystal chemistry principles. With a set of 91 experimentally determined data collected on hollandite that is available in open literature, we trained the network and subsequently tested the predictive power of the trained network. Relying on the successful outcomes of the trained network at the testing phase, we further utilized the trained network to map the dependence of the tunnel size, which was used as a criterion for Cs compatibility in the channel, in a wide compositional space encompassing eighteen 3+ cations and fifteen 4+ cations. By combining the Cs compatibility and the structural tolerance factor for hollandite structure, the predicted B-site compositions, comprising of cations spanning across the depth and breadth of the periodic table, can be employed as a guide in the search for optimal hollandite composition for Cs immobilization.