Machine learning guided discovery of potassium-selective porous silicates

Objective

CLEANSE replaces the time-consuming, expensive, and error-prone trial-and-error approach for developing cation-selective sorbents with a predictive, iterative computation-to-experiment loop. First, we will resolve why sodium zirconium cyclosilicate (ZS-9) is highly selective for K+ by training equivariant machine-learning interatomic potentials (MLIPs) on ab initio data and running long (biased) molecular-dynamics simulations under realistic aqueous, multi-ion conditions. These models will also predict experimental observables such as solid-state NMR tensor and Born effective charges for effective NMR and IR spectral prediction. Second, we will perform high-throughput screening of broad synthesizable porous silicate databases, ranking candidates by adsorption capacity, competitive selectivity against Na+/NH4+/H3O+, and diffusion barriers to derive design rules for effective/selective K+ capture. The best candidates will be synthesised and evaluated in laboratory assays, then characterised by solid-state NMR, X-ray diffraction, vibrational spectroscopy and electron microscopy. Experimental results will feed back to refine the models and optimise composition and pore architecture. CLEANSE ensures two-way knowledge transfer between my expertise in advanced simulation and the host’s strengths in synthesis and advanced characterisation, and will generate open, reusable datasets, models and workflows. Outcomes enable data-driven design of ion-exchange materials across water treatment, resource recovery, food and medical applications, while advancing fundamental understanding of selective ion capture in porous silicates.