Crystal structure prediction with nuclear quantum and finite-temperature effects via deep free energy learning
Xiaoyang Wang, Yinan Wang, Wenbo Zhao, Hanyu Liu, Hao Xie, Lei Wang, Han Wang · 2026
Read Crystal structure prediction with nuclear quantum and finite-temperature effects via deep free energy learning by Xiaoyang Wang, Yinan Wang, Wenbo Zhao, Hanyu Liu, Hao Xie, Lei Wang, Han Wang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-20230.txt.
Introduction
Wang and six colleagues predict crystal structures with heat and nuclear quantum effects, via deep free energy learning. The harmonic free energy surface shares the shape of a potential energy surface, so a deep neural net potential learns it directly. A two-level concurrent learning workflow builds the model, giving energies, forces, and stresses in one forward pass. At 200 GPa and 300 K it matches seen LaH10 and LaSc2H24 phases and finds a new stable clathrate hydride, LaScH8. Its cost runs more than a million fold below DFT-level rivals.
Catalog record
| CREATOR | Xiaoyang Wang, Yinan Wang, Wenbo Zhao, Hanyu Liu, Hao Xie, Lei Wang, Han Wang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 44609 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.20230
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Crystal structure prediction with nuclear quantum and finite-temperature effects via deep free energy learning Xiaoyang Wang, Yinan Wang, Wenbo Zhao, Hanyu Liu, Hao Xie, Lei Wang, Han Wang Abstract Accurate crystal structure prediction (CSP) requires accounting for finite-temperature and nuclear quantum effects, yet first-principles evaluation of the free energy surface (FES) remains prohibitive for high-throughput searches. We observe that the self-consistent harmonic approximation (SCHA) FES, as a function of nuclear centroid positions, shares the same mathematical structure as a potential-energy surface and can therefore be directly learned by a deep neural network potential. The resulting deep free energy (DF) model, constructed via a two-level concurrent-learning workflow, evaluates free energies, forces, and stresses in a single forward pass. Applied to the La–Sc–H system at 200 GPa and 300 K, DF-based CSP reproduces the stability of the experimentally observed LaH 10 \text{LaH}{\vphantom{\text{X}}}_{\vphantom{\text{2}}\smash[t]{\text{10}}}^{\vphantom{\smash[t]{\text{2}}}\hphantom{\text{10}}\text{}} and LaSc 2 H 24 \text{LaSc}{\vphantom{\text{X}}}_{\vphantom{\text{2}}\smash[t]{\text{2}}}^{\vphantom{\smash[t]{\text{2}}}\hphantom{\text{2}}\text{}}\text{H}{\vphantom{\text{X}}}_{\vphantom{\text{2}}\smash[t]{\text{24}}}^{\vphantom{\smash[t]{\text{2}}}\hphantom{\text{24}}\text{}} , and discovers an unreported thermodynamically stable clathrate hydride: P 4 / m m m P4/mmm LaScH 8 \text{LaScH}{\vphantom{\text{X}}}_{\vphantom{\text{2}}\smash[t]{\text{8}}}^{\vphantom{\smash[t]{\text{2}}}\hphantom{\text{8}}\text{}} . Benchmarked on the LaH 10 \text{LaH}{\vphantom{\text{X}}}_{\vphantom{\text{2}}\smash[t]{\text{10}}}^{\vphantom{\smash[t]{\text{2}}}\hphantom{\text{10}}\text{}} system, the DF model achieves a 1.72 × 10 6 1.72\times 10^{6} -fold cost reduction relative to DFT-level SSCHA. The DF framework provides a scalable route for incorporating finite-temperature and nuclear quantum effects into high-throughput crystal structure prediction. I Introduction Crystal structure prediction (CSP) has become an indispensable tool for computational materials discovery, providing promising candidates for experimental synthesis and substantially reducing the trial-and-error effort inherent in traditional approaches 40 . CSP methods have facilitated breakthroughs across diverse domains, including high-temperature superconductivity 32 , thermoelectrics 26 , superhard materials 20 , and geoscience 15 . The primary goal of CSP is to identify the global minimum of the high-dimensional classical potential-energy surface (PES) that defines the zero-temperature stability of a crystal, without accounting for nuclear quantum effects. A variety of global optimization strategies have been developed for this purpose, including simulated annealing 18 , evolutionary algorithms 23 , basin hopping 35 , random sampling 27 , and particle swarm optimization 8 , and have been implemented in widely used packages such as AIRSS 22 , USPEX 13 , MAGUS 36 , and CALYPSO 41 . Combined with first-principles calculations, these methods have made CSP a routine tool for computational materials discovery. A key gap between current CSP methods and real-world materials discovery is that predictions are based on the classical PES, which does not capture finite-temperature effects (FTE) or nuclear quantum effects (NQE) that may significantly contribute to the stability of crystal structures. For instance, \beta -Ti, a high-temperature body-centered-cubic (BCC) structure 33 , is found to be dynamically unstable in the BCC configuration when optimized using standard density functional theory (DFT) methods 31 . Another example is the ternary hydride P6/mmm \text{LaSc}{\vphantom{\text{X}}}_{\vphantom{\text{2}}\smash[t]{\text{2}}}^{\vphantom{\smash[t]{\text{2}}}\hphantom{\text{2}}\text{}}\text{H}{\vphantom{\text{X}}}_{\vphantom{\text{2}}\smash[t]{\text{24}}}^{\vphantom{\smash[t]{\text{2}}}\hphantom{\text{24}}\text{}} , recently demonstrated experimentally to exhibit room-temperature stability and superconductivity with an onset temperature of 271-298 kelvin at 195-266 GPa 30 . However, this phase is dynamically unstable on the classical PES at 200 GPa. These examples demonstrate that CSP relying solely on the classical PES would miss important structures stabilized by FTE and NQE. Modeling crystal stability with FTE and NQE requires the construction of the free-energy surface (FES), and the corresponding CSP task reduces to identifying the local minima on the FES. The precise computation of the FES is nontrivial, as it requires integration of the partition function over the high-dimensional configuration space. The simplest approximation to the FES is provided by the harmonic approximation, which expands the crystal potential energy to second order in atomic displacements 4 . To go beyond this limit, several systematic extensions have been developed, including the quasi-harmonic approximation, which incorporates volume-dependent free-energy contributions 7 , and perturbative anharmonic approaches that expand the PES to higher orders beyond the quadratic term 6 . Alternatively, the FES can be estimated from molecular dynamics (MD) simulations, either by extracting temperature-dependent effective force constants 14 or via thermodynamic integration 10 ; 34 . To explicitly incorporate NQE in addition to FTE, classical MD can be replaced by path-integral molecular dynamics (PIMD) 5 . Beyond perturbative treatments, non-perturbative frameworks such as the self-consistent harmonic approximation (SCHA) 16 and its modern stochastic formulation, the stochastic self-consistent harmonic approximation (SSCHA) 9 ; 3 ; 2 ; 25 , provide a variational description of anharmonic and quantum nuclear effects. Standard SSCHA implementations rely on first-principles methods, typically DFT, to evaluate the potential-energy contribution to the FES. More recently, these approaches have been significantly accelerated by replacing computationally intensive DFT calculations with machine-learning interatomic potentials (MLIPs) 1 ; 11 ; 28 . Despite rapid advances in FES methodologies, direct FES evaluation remains computationally demanding, and high-throughput CSP on the FES is still in its early stages. Many existing approaches conduct structure searches on the PES and subsequently evaluate the finite-temperature stability of selected candidates in a post hoc manner using thermodynamic integration 38 or solid–liquid phase coexistence molecular dynamics 37 . Kruglov et al. 19 proposed the T-USPEX framework, which incorporates free-energy evaluations at each CSP iteration; however, candidate structures are still generated via PES-based searches. More recently, MLIP-accelerated SSCHA has been integrated directly into evolutionary CSP workflows 1 ; 11 ; 28 , enabling structure relaxation on the anharmonic FES at a fraction of the first-principles cost. Nevertheless, these approaches still require explicit SSCHA free-energy minimizations, each involving hundreds of MLIP evaluations, for every candidate structure at every CSP iteration, which limits the throughput of the structure search. In this work, we observe that the SCHA free-energy surface, as a function of centroid positions alone, shares the same mathematical structure as a potential-energy surface, and can therefore be directly learned by a deep neural network potential architecture. We term the resulting model the deep free energy (DF). The DF model is constructed via a two-level concurrent-learning workflow: the first level generates a deep-learning potential-energy model (DP) as a surrogate for computationally expensive DFT calculations, while the second level constructs a DF model that approximates the SCHA free-energy surface. Once trained, the DF model evaluates the free energy, free-energy forces, and free-energy stresses in a single forward pass, entirely bypassing the stochastic SSCHA minimization required by existing MLIP-based approaches. CSP is then performed directly on this DF-approximated FES to identify structures that are stable with respect to free-energy minimization. The proposed framework is applied to the La–Sc–H ternary system, where a previously unreported thermodynamically stable clathrate hydride P4/mmm \text{LaScH}{\vphantom{\text{X}}}_{\vphantom{\text{2}}\smash[t]{\text{8}}}^{\vphantom{\smash[t]{\text{2}}}\hphantom{\text{8}}\text{}} is discovered. The remainder of this paper is organized as follows. Section II introduces the self-consistent harmonic approximation (SCHA) for the free-energy surface, its stochastic formulation (SSCHA), and the proposed deep-learning free-energy modeling workflow. Section III presents the application of the DF method to the La–Sc–H ternary system, demonstrating that the approach not only rediscovers the well-known P6/mmm \text{LaSc}{\vphantom{\text{X}}}_{\vphantom{\text{2}}\smash[t]{\text{2}}}^{\vphantom{\smash[t]{\text{2}}}\hphantom{\text{2}}\text{}}\text{H}{\vphantom{\text{X}}}_{\vphantom{\text{2}}\smash[t]{\text{24}}}^{\vphantom{\smash[t]{\text{2}}}\hphantom{\text{24}}\text{}} but also reveals a previously unreported structure. Conclusions and discussion are provided in Section IV . II Method II.1 Self-consistent harmonic approximation The self-consistent harmonic approximation (SCHA) applies to both molecular systems under open boundary conditions and crystalline systems under periodic boundary conditions. For clarity, the formulation below is presented at the \Gamma point only; extension to multiple Brillouin-zone sampling points is straightforward. We consider an atomistic system composed of N nuclei, whose quantum