Design of Ligands for Selective Separation of Ga in the Presence of Fe —An ML-DFT Approach

Gallium (Ga) is a critical strategic element indispensable in advanced technologies, including semiconductors, optoelectronics, and photovoltaic systems. Due to its limited natural occurrence and escalating demand, Ga has been designated as a critical material, underscoring its significance and supply chain vulnerabilities. Ga is predominantly recovered as a byproduct from bauxite and sphalerite ores, though its co-occurrence with iron (Fe) in these sources complicates efficient extraction due to Fe's higher abundance. Developing ligands with high selectivity for Ga over Fe is essential to optimizing separation/recovery processes, minimizing operational costs, and ensuring a stable supply of this element for critical technological applications. However, exploring a wide range of possible ligand space with stand-alone experiments is expensive and time-consuming. As a way out, we have developed a joint machine learning and density functional-based workflow to predict the metal-ligand stability constants (K) for the Ga and Fe. In doing so, we have developed a Gaussian process regression (GPR) model trained on the NIST database for predicting log K. This ML model is then used to predict M-L stability constants of unknown ligands taken from the PubChem database. Alongside this, DFT-based exploration of stability constants is also performed to validate these ML-based results. Overall, this forward model workflow accelerates the screening of existing ligands for selective Ga binding, for which the experimental stability constant values are unknown. While the forward model is used to screen the existing ligands, a supervised variational autoencoder (SVAE) is employed for the inverse design of new Ligand metal complexes. Incorporating stability constant (log K) values during the training of the VAE created a biased latent space, which is further used to sample new metal-ligand complexes and predict their stability constant. This developed generative inverse design approach allows us to automate the generation of novel ligands. As such, these developed workflows accelerated the ligand screening and the new ligand design.