Dataset for: Out-of-Distribution Robust Active Learning of Message-Passing Potentials for Extreme Non-Equilibrium Iron Dynamics

The simulation of extreme non-equilibrium materials phenomena, such as shock compression and extreme shear in body-centered cubic (BCC) iron, demands the quantum-mechanical fidelity of \textit{ab initio} methods at lengths and timescales accessible only to empirical potentials. Machine-Learned Interatomic Potentials (ML-IPs), particularly E(3)-equivariant networks like MACE, bridge this gap but suffer from catastrophic out-of-distribution (OOD) failure when encountering extreme configurations unrepresented in their training data. Furthermore, retraining ML-IPs on extreme high-temperature configurations typically induces catastrophic forgetting of 0~K harmonic equilibrium properties. Here, we introduce the Out-of-Distribution Robust Active Learning (OODR-AL) framework, a closed-loop pipeline that couples deep ensemble force-variance with Farthest Point Sampling in latent descriptor space to systematically chart extreme phase spaces without redundant sampling. Applied to BCC iron subjected to 1800~K thermal shocks and massive strain-rate shear deformations, OODR-AL converges an exceptionally robust MACE potential using only $\sim$300 high-precision Quantum ESPRESSO labels (PAW PBE, 50/400~Ry cutoff). Crucially, we demonstrate that by explicitly incorporating virial stress learning with optimized weights, our MACE potential achieves \textit{Comprehensive Fidelity}---preventing simulation collapse at extreme non-equilibrium states while strictly preserving 0~K macroscopic mechanical properties, such as elastic constants and the equation of state. This work establishes a fundamental paradigm shift in ML-IP development, proving that resilience in extreme dynamics does not necessitate the sacrifice of ground-state thermodynamic accuracy.