Data-efficient machine-learning of complex Fe-Mo intermetallics using domain knowledge of chemistry and crystallography

Atomistic simulations of multi-component systems require accurate descriptions of interatomic interactions to resolve details in the energy of competing phases. A particularly challenging case are topologically close-packed (TCP) phases with close energetic competition of numerous different site occupations even in binary systems like Fe-Mo. In this work, machine learning (ML) models are presented that overcome this challenge by using features with domain knowledge of chemistry and crystallography. The resulting data-efficient ML models need only a small set of training data of simple TCP phases $A$15, $\sigma$, $\chi$, $\mu$, $C$14, $C$15, $C$36 with 2-5 WS to reach robust and accurate predictions for the complex TCP phases $R$, $M$, $P$, $\delta$ with 11-14 WS. Several ML models with kernel-ridge regression, multi-layer perceptrons, and random forests, are trained on less than 300 DFT calculations for the simple TCP phases in Fe-Mo. The performance of these ML models is shown to improve systematically with increased utilization of domain knowledge. The convex hulls of the $R$, $M$, $P$ and $\delta$ phase in the Fe-Mo system are predicted with uncertainties of 20-25 meV/atom and show very good agreement with DFT verification. Complementary X-ray diffraction experiments and Rietveld analysis are carried out for an Fe-Mo R-phase sample. The measured WS occupancy is in excellent agreement with the predictions of our ML model using the Bragg Williams approximation at the same temperature.