Abstract: Modeling oxygen interstitials in titanium requires a new empirical potential. We optimize potential parameters using a fitting database of first-principle oxygen interstitial energies and forces. A new database optimization algorithm based on Bayesian sampling is applied to obtain an optimal potential for a specific testing set of density functional data. A parallel genetic algorithm minimizes the sum of logistic function evaluations of the testing set predictions. We test the transferability of the potential model against oxygen interstitials in HCP titanium, transition barriers between oxygen interstitial sites, and oxygen in the titanium prismatic stacking fault. The potential predicts that the interaction between oxygen and a screw dislocation core is weak and short-ranged.