The DeepMT model of Tao, Wang, Liu, and Lian (Angew. Chem. Int. Ed. 2025) combines a chemical-potential-based mesoscopic thermodynamic description of electrolytes — in practice a classical density functional theory (CDFT) of the electric double layer — with a deep neural network surrogate, predicting equilibrium ion density profiles roughly four orders of magnitude faster than direct theoretical calculation while reproducing ion adsorption, surface charge, and differential capacitance. Its validation envelope, however, covers only one-dimensional planar electrodes, equal-magnitude monovalent ions, hard-sphere plus Coulomb plus correlation physics, a fixed temperature, and equilibrium states. The open problem is to carry this learned-operator strategy to realistic electrochemical interfaces: three-dimensional and heterogeneous or reactive electrode geometries, multicomponent and mixed-valence electrolytes with explicit solvent structure and additional interactions, temperature variation, time-dependent (dynamic) charging, reliable transfer across CDFT, molecular dynamics (MD), and experimental data, and integration into coupled transport–reaction simulations with experimental validation of downstream observables. As of August 2026 only the explicit-solvent strand has begun to be addressed (the same group's explicit-solvent follow-up ES-DeepMT, Chem. Eng. Sci. 2026), three-dimensional learned classical functionals exist only for non-electrolyte simple and lattice fluids, and no learned electric-double-layer operator has been extended to realistic electrodes or coupled into multiphysics simulation; the remaining core is open.

Background

The electric double layer (EDL) is the nanometer-scale interfacial region in which ions of an electrolyte reorganize near a charged electrode, screening its charge in layered fashion. EDL structure controls interfacial charge storage, adsorption, and reaction rates, and therefore the performance of batteries, supercapacitors, and electrocatalysts. Theoretical tools for it form a well-known accuracy-versus-speed hierarchy. Mean-field Poisson–Boltzmann theory describes the electrolyte as point ions in a dielectric continuum and is cheap, but it neglects ion size, short-range correlations, and the oscillatory layering of dense electrolytes, and fails for concentrated or strongly interacting solutions. Molecular dynamics (MD) and Monte Carlo simulations capture microscopic structure but are too expensive for wide parameter sweeps or device-scale coupling. Classical density functional theory (CDFT) is an intermediate, mesoscopic equilibrium route: the equilibrium density profiles follow from minimizing a free-energy functional that can include hard-sphere exclusion, Coulomb interactions, and correlation contributions; in practice each parameter point requires iterative solution of many coupled equations. The source paper works with a restricted primitive model electrolyte (ions modeled as charged hard spheres of equal diameter and opposite unit charge in a structureless dielectric solvent) at a planar electrode. Its DeepMT model (deep neural network enhanced mesoscopic thermodynamics) trains a neural network on CDFT data to directly map thermodynamic conditions — bulk concentration, electrode potential, ion size, dielectric constant — to equilibrium ion density profiles, from which interfacial observables such as ion adsorption, surface charge density σ\sigma, and differential capacitance Cd=∂σ/∂UC_d=\partial\sigma/\partial U are obtained by statistical analysis; the surrogate is about four orders of magnitude faster than the direct CDFT calculation it replaces. This makes feasible high-throughput sweeps and, prospectively, use of the EDL operator as a component inside continuum transport and reaction simulations. The same research direction is part of a rapidly growing effort to learn classical density functionals and their operators from data, including neural functional theory (Sammüller, Hermann, de las Heras, and Schmidt, PNAS 2023), machine-learned functionals for ionic fluids (Bui and Cox, Phys. Rev. Lett. 2025), neural operators for one-dimensional CDFT mappings (Pan et al. 2025), and three-dimensional learned functionals for simple and lattice fluids (Cheng 2026; Weimar, Oettel, and Simon 2026; Glitsch, Weimar, and Oettel 2025). What none of these delivers is a learned EDL operator valid under the conditions of real electrochemical interfaces — non-planar and chemically heterogeneous electrodes, multicomponent mixed-valence electrolytes with explicit solvent, non-isothermal and time-dependent operation — together with reliable cross-paradigm transfer and validated coupling into multiphysics simulation.

Problem Statement

The DeepMT framework of Tao et al. (2025) is demonstrated only for a one-dimensional planar electrode in contact with an equal-magnitude monovalent model electrolyte with hard-sphere steric, Coulomb, and correlation physics at a single fixed temperature, in equilibrium. The problem is to extend this learned mesoscopic-thermodynamic operator strategy to the conditions of real electrochemical interfaces, as one coherent objective — a fast, thermodynamically consistent surrogate of electrode/electrolyte interface structure and response usable under realistic interfacial conditions and inside multiscale simulation. Concretely, the interlinked requested extensions are: (i) validity beyond one-dimensional planar geometry, i.e. fully three-dimensional electrode geometries and heterogeneous or reactive surfaces with specific ion–surface chemistry; (ii) electrolyte complexity beyond the restricted primitive model, i.e. multicomponent and mixed-valence electrolytes, explicit solvent structure, and additional interaction terms such as van der Waals attraction, ion pairing/association, and polarizability; (iii) temperature-dependent and time-dependent (dynamic charging) operation, not only equilibrium at fixed temperature; (iv) a demonstrated, reliable strategy for operator transfer and training-data fusion across modeling paradigms (CDFT, MD, and experiment) and for generalization beyond the trained parameter domain; and (v) integration of the learned EDL operator into coupled multiphysics simulations — ion transport, heat, fluid flow, and surface reactions — with the resulting downstream observables validated against experiments and high-fidelity simulations under complex operating conditions. Within each extension the listed features describe its scope rather than separate deliverables, and the five extensions are independent, each defined by its full scope rather than by an isolated feature — in particular, explicit solvent structure alone, already treated in ref. 2, does not by itself constitute (ii) — and validated in-domain against the appropriate reference paradigm (CDFT, molecular dynamics, or experiment), while the extensions not addressed remain open.

References

  1. Tao, Haolan; Wang, Sijie; Liu, Honglai; Lian, Cheng. Deep Neural Network Enhanced Mesoscopic Thermodynamic Model for Unlocking the Electrode/Electrolyte Interface. Angewandte Chemie International Edition 64(6), e202418447 (2025). DOI: 10.1002/anie.202418447. https://doi.org/10.1002/anie.202418447
  2. Zhu, Mingyang; Wang, Sijie; Lin, Yiting; Liu, Honglai; Lian, Cheng. Modeling the solvent effect on electric double layers using machine learning enhanced mesoscopic thermodynamics model. Chemical Engineering Science 329, 123825 (2026). DOI: 10.1016/j.ces.2026.123825. https://doi.org/10.1016/j.ces.2026.123825
  3. Bui, Anna T.; Cox, Stephen J. Learning Classical Density Functionals for Ionic Fluids. Physical Review Letters 134, 148001 (2025). DOI: 10.1103/PhysRevLett.134.148001. https://doi.org/10.1103/PhysRevLett.134.148001
  4. Sammüller, Florian; Hermann, Sophie; de las Heras, Daniel; Schmidt, Matthias. Neural functional theory for inhomogeneous fluids: Fundamentals and applications. Proceedings of the National Academy of Sciences 120(50), e2312484120 (2023). DOI: 10.1073/pnas.2312484120. https://doi.org/10.1073/pnas.2312484120
  5. Cheng, Bingqing. Equivariant learning of a transferable three-dimensional classical density functional. arXiv:2608.13506 (2026). https://arxiv.org/abs/2608.13506
  6. Weimar, Jens; Oettel, Martin; Simon, Alessandro. Cubic-Equivariant Neural Density Functional Theory for Three-Dimensional Lattice Fluids. arXiv:2608.08137 (2026). https://arxiv.org/abs/2608.08137
  7. Glitsch, Felix; Weimar, Jens; Oettel, Martin. Neural Density Functional Theory in Higher Dimensions with Convolutional Layers. arXiv:2502.13717 (2025). https://arxiv.org/abs/2502.13717
  8. Pan, Runtong; Fang, Xinyi; Azizzadenesheli, Kamyar; Liu-Schiaffini, Miguel; Gu, Mengyang; Wu, Jianzhong. Neural Operators for Forward and Inverse Potential-Density Mappings in Classical Density Functional Theory. arXiv:2506.06623 (2025). https://arxiv.org/abs/2506.06623
  9. Cao, Zhonglin; Wang, Yuyang; Lorsung, Cooper; Farimani, Amir Barati. Neural Network Predicts Ion Concentration Profiles under Nanoconfinement. arXiv:2304.04896 (2023). https://arxiv.org/abs/2304.04896
  10. Huang, Pan; Leng, Yifei; Lian, Cheng; Liu, Honglai. Porous-DeepONet: Learning the Solution Operators of Parametric Reactive Transport Equations in Porous Media. Engineering 39, 94–103 (2024). DOI: 10.1016/j.eng.2024.07.002. https://doi.org/10.1016/j.eng.2024.07.002
  11. Ma, Ke; Janssen, Mathijs; Lian, Cheng; van Roij, René. Dynamic density functional theory for the charging of electric double layer capacitors. The Journal of Chemical Physics 156, 084101 (2022). DOI: 10.1063/5.0081827. https://doi.org/10.1063/5.0081827
  12. Shi, Jiawei; Tian, Yun; Sun, Mengwei; Zhang, Xu; Zhou, Zhen. Artificial intelligence-facilitated multiscale theoretical modelling to advance battery science. Science China Chemistry 69(6), 2821–2835 (2026). DOI: 10.1007/s11426-025-3306-8. https://doi.org/10.1007/s11426-025-3306-8
  13. Simon, Alessandro; Oettel, Martin. Machine Learning Approaches to Classical Density Functional Theory. In: Machine Intelligence for Materials Science (2026). DOI: 10.1007/978-3-032-04129-6_6. https://doi.org/10.1007/978-3-032-04129-6_6