Quantitative Mapping of Oxygen Affinity to Local Chemical Environments in Ti–Zr–Nb–Ta Alloys via Machine Learning

Simultaneously enhancing strength and ductility is a longstanding challenge in materials science. Recent studies show that incorporating oxygen into Ti–Zr−family refractory high-entropy alloys (HEAs) can overcome this trade-off, with improved properties stemming from interstitial oxygen occupancy. However, oxygen occupancy is inherently site-specific and strongly influenced by local chemical environments, complicating quantitative predictions of oxygen solution energies at individual sites. Here, we address this challenge in the Ti–Zr–Nb−Ta system by combining high-throughput first-principles calculations with machine learning (ML). Representing local environments with Smooth Overlap of Atomic Positions features, our ML model accurately predicts oxygen solution energies from initial, unrelaxed atomic configurations (R2 = 0.93, mean absolute error = 0.11 eV), enabling analysis of oxygen occupancy trends and spatial correlations over extensive compositional ranges. Two critical descriptors─the average oxygen solution energy and its standard deviation─are proposed to quantify overall oxygen affinity and distribution heterogeneity within each composition. Notably, these descriptors correlate closely with experimentally reported strength and ductility enhancements, highlighting that controlled oxygen interstitial occupancy is crucial for optimizing mechanical properties. Our findings provide fundamental insights into oxygen solution behaviors in HEAs and facilitate the design of oxygen-containing HEAs with controlled oxygen incorporation and distribution.