Shivansh Raj

LinkedIn Profile

This project applied machine learning to accelerate the discovery of new dielectric materials for next-generation high voltage power grid insulation. Current electrical systems rely on sulphur hexafluoride (SF6) gas as an insulator, which is a potent greenhouse gas 23,000 times more harmful than CO2. Replacing it requires identifying solid materials with large band gaps, the energy barrier that prevents electricity from flowing through a material.

Data from 4,766 hafnium oxide and zinc oxide based materials was downloaded, cleaned and filtered to 2,076 high quality training examples, each described by 165 compositional features. The key innovation was grouping materials by their crystal space group symmetry, specifically isolating the P1 space group before training dedicated machine learning models on each group.

The best model, XGBoost trained on P1 materials from both systems combined, achieved an R² of 0.845 and a mean absolute error of 0.269 eV. Screening 685 unseen candidate materials identified 55 with predicted band gaps above the 4 eV threshold for high voltage insulation, with the top candidate RbHfC(OF)4 predicted at 5.94 eV. The model was further validated on aluminium oxide and tin oxide systems, achieving an R² of 0.55 on aluminium oxide. All computation was performed on the Barkla2 HPC cluster at the University of Liverpool.

What were the key results of your research project?

• Best model: XGBoost trained on P1 space group materials from both HfO and ZnO systems combined, achieving R² = 0.845 and MAE = 0.269 eV
• Space group number was identified as the most important predictor of band gap, accounting for 14% of feature importance — validating the decision to group materials by crystal symmetry
• 685 unseen candidate materials were screened, with 55 predicted above the 4 eV threshold required for high voltage insulation
• Top candidate: RbHfC(OF)4 with a predicted band gap of 5.94 eV — fluorine-containing compounds consistently dominated the top candidates
• The model generalised well to aluminium oxide (R² = 0.55) confirming transfer of chemical knowledge across oxide systems, though tin oxide proved too chemically different (R² = 0.04)
• Grouping by P1 space group improved HfO model performance from R² = 0.720 (ungrouped) to R² = 0.808 (P1 only), and combining both systems further improved it to R² = 0.845

A presentation of this research will be shared here and on our YouTube site when available.