Uncertainty-aware machine learning framework for predicting dislocation plasticity and stress–strain response in metallic alloys, Part I : FCC systems

The discovery of alloys with superior mechanical properties is hindered by the inability of existing predictive models to be fast, transferable, and uncertainty aware simultaneously. On the one hand, conventional crystal plasticity methods are computationally expensive and commonly rely on phenomenological laws that require experiment-specific calibration. A classical example is the Kocks–Mecking–Estrin (KME) model for the evolution of dislocation density as a function of strain and grain size, which suffers from poor generalization across even the same material with different microstructural features. On the other hand, deterministic machine learning frameworks, while fast, overlook substantial uncertainties in experimental data. Here, we present a physics-informed, uncertainty-aware framework for face-centered cubic (FCC) alloys that combines dislocation physics with machine learning. A mixture density network, trained on literature stress–strain data for polycrystalline Ni, Cu, Al, and stainless steels, predicts the probability distributions of the dislocation density evolution. These distributions are mapped to stress and upscaled through stochastic homogenization to output confidence bounds that capture experimental scatter. Without recalibration, the framework successfully extends beyond its training data to multicomponent FCC alloys (NiCoCr and NiCoCrMnFe) through physics-based parameter adjustments alone. This approach enables mechanism-aware uncertainty quantification and reliable, high-throughput screening of FCC alloys, serving as a fast and accurate drop-in surrogate for higher-fidelity models.