Session: 04-06-01: AI for Heterogeneous Materials Design, Discovery, and Manufacturing I

Paper Number: 173566

A Data-Driven Statistical Approach to Predicting Stress-Strain Curves and Dislocation Evolution in Fcc and Bcc Alloys With Quantified Uncertainty 

The rapid advancement of autonomous alloy discovery critically depends on the development of efficient, physics-consistent models that can accurately predict mechanical properties such as stress–strain behavior in complex polycrystalline systems. Traditional empirical and phenomenological approaches are often limited by oversimplification and poor transferability to different material systems, especially in the presence of significant microstructural variability and incomplete experimental data. In this work, we propose a hybrid data-driven and physics-based framework to predict the evolution of dislocation-mediated plasticity and the resulting mechanical response in novel high-cost alloys, incorporating uncertainty explicitly.

Focusing first on face-centered cubic (FCC) metals and alloys, we construct a robust predictive pipeline based on a mixture density network (MDN) trained on a curated dataset of experimental stress–strain curves. The MDN predicts the full probability distribution of dislocation density as a latent variable, conditioned on microstructural and loading descriptors such as strain, grain size, strain rate, and initial dislocation content. These predictions are propagated through a generalized Taylor hardening law to compute grain-level stress distributions, which are then homogenized using a parallel–series model to yield macroscopic stress–strain curves with well-calibrated uncertainty bounds.

A distinctive aspect of this approach is its treatment of microstructure as a stochastic field, allowing us to model polycrystals using synthetic representative volumes with statistically distributed grain sizes, orientations, and dislocation densities. This facilitates efficient sampling of stress–strain responses under varying processing conditions without the need for full-field simulations or phenomenological curve fitting. The framework captures both aleatory uncertainty, arising from inherent microstructural variability, and epistemic uncertainty, introduced by inconsistent or incomplete experimental data. By comparing model predictions to experimental observations across a diverse set of FCC materials, including pure metals and complex multicomponent alloys, we show that our framework achieves high predictive accuracy without per-material-system recalibration. It also quantitatively identifies which microstructural features contribute most to response variability, guiding more informative data collection in future experiments.

Moreover, the model is designed for seamless integration with high-throughput experimental platforms and autonomous laboratory systems. Its probabilistic output format is compatible with Bayesian optimization and other decision-making algorithms commonly used in closed-loop discovery frameworks. When experimental data deviate significantly from predicted uncertainty bounds, the model can be used diagnostically to flag potentially unmeasured or misreported features, such as residual stress states, texture, or local chemical heterogeneity, offering a systematic way to improve data fidelity and interpretability.

Building on the demonstrated success in FCC systems, ongoing work focuses on extending the framework to body-centered cubic (BCC) polycrystals, which present additional modeling challenges due to non-Schmid effects, stronger strain-rate sensitivity, and thermally activated dislocation glide. The BCC extension includes mechanism-specific evolution laws and data-driven learning of activation parameters under uncertainty. We show that our framework is a transferable, mechanism-aware, and uncertainty-aware model for predicting the stress-strain response of polycrystalline alloys, which can then be seamlessly interfaced with high-throughput synthesis, characterization, and design.

Presenting Author: Jaafar El-Awady Johns Hopkins University

Presenting Author Biography: Jaafar El-Awady is a professor of mechanical engineering at Johns Hopkins University. He directs the Integrated Computations and Experiments for Intelligent Materials Design Laboratory (ICE-4-iMD), which focuses on advancing the understanding of material deformation in extreme environments. His lab's work focuses on predicting the mechanical behavior and failure of complex alloys, coatings, polymers, and fatigue-prone materials through integrating machine learning techniques, multi-scale computational modeling, and experiments to develop next-generation materials for aerospace, naval, automotive, and energy applications.

El-Awady has earned numerous honors, including the NSF CAREER Award, DARPA Young Faculty Award, ASME Orr Early Career Award, and the TMS Brimacombe Medal. Before joining Hopkins in 2010, he worked with the Air Force Research Laboratory and Digital Material Solutions Inc. He holds degrees in aerospace engineering from Cairo University and a PhD from UCLA.

Authors:

Jing Luo Johns Hopkins University
Yejun Gu Johns Hopkins University
Jaafar El-Awady Johns Hopkins University