The Impact of Domain Knowledge on Universal Models for Predicting High Entropy Materials

The cornerstone of accelerated materials discovery is the development of predictive models or theories that capture stability across a wide chemical space. Although multiple foundational AI models have emerged for materials discovery, the effectiveness of adapting these models to specific material research domains remains uncertain. In this talk, we will present our data mining efforts across two high-entropy materials datasets, which include 126,429 high-entropy alloys and 18,810 high-entropy disordered rocksalt oxides and oxyfluorides computed via density functional theory. Our findings indicate that large datasets and deep graph neural frameworks incorporating many-body interactions are not always effective unless chemical trends are considered during sampling. We will also demonstrate that purely physical models can, in some cases, outperform deep graph neural networks in predicting the stability of high-entropy materials.