Ephemeral Data Derived Potentials

• Ephemeral Data Derived Potentials are machine-learned force fields dynamically constructed during structure searches using a compact, problem-specific ab initio dataset.
• They iteratively train on local configurations to rapidly evaluate energies, forces, and stress tensors, significantly reducing computational cost for complex systems.
• EDDPs accelerate global optimization and metastability screening in superhydride and hydride superconductor discovery, guiding the prediction of key superconducting properties.

Ephemeral Data Derived Potentials (EDDPs) constitute a class of machine-learned interatomic potential energy surfaces constructed on-the-fly and targeted toward the efficient first-principles screening of crystal structure landscapes, particularly under extreme conditions such as high pressure. EDDPs combine a limited set of ab initio reference calculations with rapid, iterative local fitting to enable accelerated global optimization, crystal structure prediction, and stability analysis for chemically complex and data-starved composition spaces—most notably superhydride and hydride superconductors at multi-gigapascal pressures (Caussé et al., 22 Dec 2025).

1. Definition and Theoretical Basis

Ephemeral Data Derived Potentials are formed dynamically during structural searches by repeatedly training machine-learning or regression-based force fields using a compact, problem-specific data set generated on demand from quantum mechanical (e.g., DFT) calculations. In contrast to transferable interatomic potentials, EDDPs are “ephemeral”—valid solely for the manifold of local configurations probed during a single structure search, and discarded afterward. This approach circumvents the need for extensive, transferable training data and seeks instead high accuracy along the optimization path (Caussé et al., 22 Dec 2025).

EDDPs operate within a multiscale framework: initial candidate structures are generated, their local energy landscapes mapped via EDDPs, and promising minima refined by full quantum mechanical relaxation. The key innovation lies in the use of in situ potential fitting to rapidly evaluate energies, forces, and, when appropriate, stress tensors for large supercells and complex compositions beyond what is feasible with direct DFT sampling.

2. Methodological Implementation

The EDDP protocol involves the following systematic workflow:

1. Initial Sampling: Generate initial structures, either randomly or seeded, covering the relevant stoichiometry and pressure range.
2. Reference Calculation: Compute single-point energies and forces for these structures using a high-level electronic structure method (e.g., DFT with PBEsol exchange-correlation).
3. Local Fit: Train an EDDP (such as a moment tensor potential, Gaussian approximation potential, or neural network) on this ephemeral dataset.
4. Structure Exploration: Use the EDDP to drive local/global optimizations, molecular dynamics, phonon spectra, or convex hull construction, with rapid feedback.
5. Adaptive Refinement: As the search identifies novel low-energy configurations, select new reference points for direct DFT evaluation, iteratively refine the EDDP, and converge the search.

EDDPs are discarded once the search concludes; they are not intended for transfer to unrelated systems or conditions. Their application is tailored for high-throughput, automated pipelines, particularly in compositionally complex multinary hydrides where the combinatorial phase space is prohibitive for direct ab initio enumeration (Caussé et al., 22 Dec 2025).

3. Applications in Superhydride and Hydride Superconductor Discovery

EDDPs have been central to the computational discovery and screening of high-pressure ternary hydrides, enabling rapid construction of phase diagrams (convex hulls), refinement of thermodynamic stability, and dynamical/metastability analysis. In the identification of A15-type YSbH₆ as a metastable hydride superconductor, EDDPs guided extensive Y–Sb–H structure searches at pressures up to 120 GPa (Caussé et al., 22 Dec 2025). By accelerating the search for local and global minima, EDDPs provided an efficient mapping of the energetic landscape, yielding the following outcomes:

• Discovery of Pm$\bar{3}$-YSbH₆ as a dynamically and kinetically stable phase at 50 GPa and 120 GPa.
• Quantification of the phase’s distance above the convex hull (e.g., 108 meV/atom at 50 GPa, 26 meV/atom at 120 GPa), comparable to known metastable oxides.
• Validation of dynamical and elastic stability through explicit phonon and elastic tensor evaluations, with molecular dynamics (driven by EDDP) confirming H-cage robustness.
• Acceleration of supercell relaxations, band structure, and phonon spectrum analysis required for accurate superconducting property prediction.

A critical benefit is the ability to efficiently filter multinary phase spaces for candidates that simultaneously satisfy stability, dynamical, and superconducting performance requirements (Caussé et al., 22 Dec 2025).

4. Technical Comparative Context

Contrasted with conventional static databases or statically trained machine-learned potentials, EDDPs are purposely non-transferable and tailored for each structure search, drastically reducing the computational cost for systems where transferable potentials do not yet exist or are not sufficiently accurate. This is particularly relevant in superhydride chemistries, where conventional empirical potentials are inapplicable. EDDPs are complementary to—rather than replacements for—direct DFT or quantum chemistry, serving to front-load the search and reserve ab initio resources for targeted final optimization and property calculations (Caussé et al., 22 Dec 2025).

5. Implications for Materials Design and Metastability Screening

EDDPs enable the systematic exploration of metastability windows for complex hydride phases at extreme conditions, translating to actionable synthesis blueprints. For example, the identification of YSbH₆ with a 26 meV/atom energetic offset from the convex hull at 120 GPa and persistent dynamical/kinetic stability at lower pressures exemplifies the role of EDDPs in guiding experiment (Caussé et al., 22 Dec 2025). The workflow provides a robust filter chain: phases are first screened thermodynamically and dynamically (using EDDP-accelerated DFT), then further evaluated for kinetic and elastic stability, electron-phonon coupling, and $T_c$ performance—thereby maximizing the probability of experimental realization in diamond anvil cell syntheses.

The competitive pressure range, manageable metastability, and clear superconducting signatures produced by this approach underpin advances in “quenchable” high-$T_c$ hydride design strategies. Future searches are expected to follow this paradigm, with EDDPs as core computational drivers for complex superhydride discovery (Caussé et al., 22 Dec 2025).

6. Limitations and Outlook

While EDDPs provide significant acceleration, their accuracy is confined to the local phase space sampled and is not guaranteed for all possible atomic environments. This inherent limitation restricts their use to rapid structure searches and stability screening, necessitating further quantum mechanical validation for ultimate property prediction and potential application design. The continued convergence of machine learning, high-throughput structure search strategies, and targeted ab initio validation—exemplified by EDDP frameworks—will remain central to hydride superconductor research, especially in the drive toward lower-pressure, quenchable, and higher-$T_c$ materials (Caussé et al., 22 Dec 2025).