Session: 13-19-02: Scientific Machine Learning (SciML) for Characterization, Modeling, and Design of Structures and Materials II

Paper Number: 173626

Uncertainty-Aware Deep Learning for Predicting Phase Stability of Complex Refractory Multi-Principal Element Alloys (Rmpeas) 

Refractory multi-principal element alloys (RMPEAs) are an emerging class of materials that exhibit exceptional mechanical properties particularly at elevated temperatures, making them promising candidates for applications in extreme environments. The complex interplay of deformation mechanisms that are chemistry and phase-dependent directly control the properties of RMPEAs. Therefore, predicting the phases that will form in each composition is a crucial first step in alloy design. However, accurately predicting phase formation in RMPEAs remains a longstanding challenge due to the enormous compositional space, non-trivial thermodynamic interactions among multiple elements, and the limited understanding of all factors influencing phase stability. In this work, we present a data-driven framework based on mixture density networks (MDNs) to predict temperature-resolved phase fractions in RMPEAs while quantifying prediction uncertainty. Our model is trained on a large dataset of CALPHAD-derived phase information covering hundreds of thousands of RMPEA compositions across a broad temperature spectrum.  The input space consists of 41 carefully selected features, chosen based on commonly used descriptors in the RMPEA community for phase prediction tasks. The model demonstrates high predictive accuracy across up to six distinct phases and a broad temperature range, while also providing uncertainty predictions that support more informed decision-making in materials design. To assess the contribution of each feature to model performance, we conduct a comprehensive feature importance and ablation study. Our results reveal that a reduced set of 12 most important features is sufficient to maintain high accuracy and low uncertainty. However, further feature removal leads to a sharp rise in uncertainty, underscoring the value of chemically meaningful descriptors for phase prediction in complex alloy systems. Beyond prediction, we integrate the MDN model into an active learning framework to accelerate materials discovery in previously unexplored regions of the composition space. By strategically targeting areas within the Ti-alloy design space where the model exhibits high uncertainty, we efficiently expand the Ti-free training dataset using a limited number of CALPHAD evaluations on Ti-containing alloys. This approach enables the discovery of novel single-phase BCC Ti-alloys. Notably, this strategy is robust with respect to the batch size of the newly added Ti-alloys, underscoring its flexibility in how new data is incorporated. Moreover, the discovered single-phase BCC compositions span diverse regions of the design space, demonstrating the model’s ability to generalize beyond the initial training distribution and uncover promising candidates in previously unexplored compositional domains. We further demonstrate the model’s utility by applying it to Ti–Cu alloys synthesized by our collaborators and by guiding the design of new single-phase BCC compositions in the Ti–Ni system, thereby successfully validating our active learning approach. Altogether, our framework provides a scalable, uncertainty-aware approach for predicting phase stability in RMPEAs and supports the targeted exploration of promising compositional regions, significantly accelerating the design of high-performance refractory alloys.

Presenting Author: Christopher Stiles Johns Hopkins University

Presenting Author Biography: Dr. Christopher Stiles is the chief scientist of the Electrical and Mechanical Engineering Group, at the Johns Hopkins Applied Physics Laboratory (JHUAPL) where he is responsible for strategic planning and execution of advanced research initiatives. He focuses on pioneering computational methods and artificial intelligence tools to facilitate scientific discovery. Dr. Stiles has an extensive publication record and research interests in topics related to multiscale modeling, machine learning, high performance computing, and computational physics. His work emphasizes convergent research between data scientists and physical scientists in areas such as quantum engineering, materials science, advanced manufacturing, chemistry, biology, and in-situ resource utilization (ISRU).

In harmony with his role at JHUAPL, Dr. Stiles is an Assistant Research Professor in the Johns Hopkins Department of Mechanical Engineering and serves as the Vice Chair of the Mechanical Engineering program within the Johns Hopkins Engineering for Professionals program. He is also a Fellow of the Hopkins Extreme Materials Institute (HEMI), contributing to interdisciplinary research efforts that push the boundaries of materials science and engineering

Authors:

Christopher Stiles Johns Hopkins University
Ali Shargh Johns Hopkins University
Jaafar El-Awady Johns Hopkins University