Within a composite cathode, the particles are randomly oriented and form pathways that may be either open or blocked to Li transport (Figure 1b). These particles also contain defects, such as interstitials, stacking faults, dislocations, or grain boundaries (GBs) that allow or hinder Li diffusion. GBs, defined as solid surfaces of contact among surface slabs of different orientation (Figure 1c), are specified by a tilt angle θ and rotation axis o. In the simplest case, when only two differently oriented grains meet, the extension of the lattice from one crystal into the other (red lines in Figure 1c) results in overlapping sites (yellow circle). The reciprocal of the fraction of all overlapping sites is quantified as Σ: this value is 1 for a perfect crystal, low for a highly symmetric GB, and high for low-symmetry GB. Altogether, with the terminating slab denoted by Miller indices (hkl), the categorization of a GB is given by the notation, Σ(hkl). The notation is convenient when describing GBs within a heterogenous, bulk material, such as those within SSEs.

If properly engineered, GBs in the SSE can play a crucial role in preventing cell failure, as they significantly influence power density, cycle life, and safety (Yu and Siegel, 2017). Due to their different local environments compared to in the bulk, GBs exhibit deviations in ionic conductivity, electronic structure, mechanical properties (Dawson, 2024). However, the diversity of GBs can be hard to resolve, let alone interpret.

To this end, polycrystalline modeling using atomistic simulations can be a precise tool for resolving buried SSE | SSE interfaces. A recent Perspective by Dawson extensively reviews the pool of GB in SSE studied so far using atomistic modeling: Li_3xLa_(2/3)–xTiO₃ (0 < x < 0.16, LLTO), Li₃OCl, Li₂OHCl, Li₇La₃Zr₂O₁₂ (LLZO), Li_0.375Sr_0.4375Ta_0.75Zr_0.25O₃ (LSTZ0.75), LiZr₂ (PO₄)₃ (LZP), Li₁₀GeP₂S₁₂, Li₃PS₄, and Li₃InCl₆ (Dawson, 2024). Therefore, we only briefly mention previous works concerning modeling of GB, focusing more on the growth in methodologies (classical approaches, ab initio force fields, and machine learning force fields) used in SSE | SSE GB modeling.

The earliest studies of SSE GB have used classical molecular dynamics (CMD) to recommend processing conditions for LLZO to increase low-energy, compact GB (Yu and Siegel, 2017), resolve origins of sluggish Na conductivity in oxide-based polycrystalline Na₃PO₄ SSE compared to sulfide-based polycrystalline Na₃PS₄ (Dawson et al., 2019), and suggest nanosizing of Li₁₀GePS₁₂ (LGPS) to increase the dimensionality of Li-ion conductivity (Dawson and Islam, 2022). The sizes of these systems range from 10³–10⁵ atoms and are deployed for around 10 nanoseconds. They are based on classical force fields (e.g., Morse potentials) that are fitted to reproduce certain static and dynamic experimental properties (e.g., lattice constants, thermal expansion, phase transitions, shear, bulk, and elastic moduli) or freshly fitted via empirical procedures (e.g., General Utility Lattice Program (Gale and Rohl, 2003)). From these works, it is evident that findings from fitted CMD can capture the local chemical and structural environments of polycrystals occupying space on the order of 10³ nm³, build experimentally consistent understandings of ion diffusion, and improve processing conditions.

CMD simulations can then inform continuum-level modeling and construct equivalent circuit models to assess how GB size controls ionic diffusion. For example, in the case of anti-perovskite Li₃OCl, four candidate GBs have remarkably low GB energies and are presumed to form with high probability during synthesis. The ease of forming these GBs could explain the discrepancy between calculated single-crystal activation barriers compared to experimental measurements, especially when grain sizes are less than a few hundred nanometers (Figure 2a). (Dawson et al., 2018) Their work emphasizes that, in addition to bulk migration barriers, grain boundary migration energies are another important parameter to consider in the search for realistic high-performance SSE. This challenge is also noted by a recent computational Review for the design of fast Li conductors (Jun et al., 2024).

While CMD is useful for obtaining microscopic descriptions of Li diffusion via a non-reactive picture of ionic transport, it captures no electronic information. However, it is known that electron conduction in SSE can lead to unwanted Li deposition and detrimental cell shorts. For example, Demir et al. use hybrid DFT and thermodynamic arguments in a 192-atom cell to show that in cubic LLZO, Li metal nucleation can only occur around regions which relax lattice strain (e.g., GBs, voids) due to the enormous positive formation energy of neutral Li formation in the bulk (Demir et al., 2024). Their investigations reveal that GBs are effective Li metal deposition sites and candidates for dendrite formation.

Understanding mixed ionic and electronic transport along SSE GB can be achieved through AIMD which uses forces from ab initio calculations and classical Newtonian mechanics to propagate equations of motion (Car and Parrinello, 1985). Quirk and Dawson use AIMD in four SSE, Li₃OCl, Li₂OHCl, β -Li₃PS₄, and Li₃InCl₆, to find that the GBs in all four materials have reduced band gaps and pronounced electron or hole traps, the latter which can form on under-coordinated anions (Figure2b). (Quirk and Dawson, 2023) Any ion at GBs that can exist in multiple states, such as S^2-/S⁻ in Li₃PS₄, O^2-/O⁻ in Li₃OCl, or In³⁺/In²⁺ in Li₃InCl₆, are thermodynamically favored centers for hole or electron polaron formation. This is problematic in terms of dendrite formation, as electron polaron localization sites are then prime locations for unwanted Li metal deposition via Li reduction: Li⁺ + In²⁺ - > Li⁰ + In³⁺. A similar observation is made in LLZO: AIMD reveals that Li-ion transport along a high-index tilt GB is remarkably high, comparable to that in the bulk, and that along this GB, there is preferential localization of excess electrons at ZrO₅ units and resultant Li metal accumulation (Gao et al., 2022). Therefore, in LLZO, the high-index GB provide culpable sites for Li metal deposition. Both studies indicate the criticality of electronic structure calculations in SSE GBs, as their electronic properties can deviate significantly from the bulk and be key predictors of failure.

While AIMD offers improved accuracy over CMD by better capturing the complex bonding expected in some high angle GBs, it remains limited to small system sizes and short timescales (e.g., typically fewer than 400 atoms and ∼100 picoseconds (ps) for Li₃OCl and Li₃InCl₆ (Quirk and Dawson, 2023), preventing the exhaustive study of large length scales needed for high angle or amorphous GBs. Additionally, as DFT approaches have systematic errors, results of AIMD simulations must sometimes undergo a secondary analysis step, using a higher-level of theory which is even more computationally expensive (Gao et al., 2022).

Studying high angle GBs at nanometer-length scales using expensive levels of theory requires going beyond CMD or AIMD approaches. ML interatomic potentials (MLIPs), which can quite accurately reproduce electronic structure calculations, have been bridging the methods gap between the accuracy of ab initio methods and relatively lower cost of classical methods since Behler and Parinello first introduced the use of high-dimensional neural network potentials in 2008 (Behler and Parrinello, 2007). The advantage of MLIPs is that they enable quantum mechanically accurate molecular simulations of systems for now typically 10³ atoms (Jacobs et al., 2025), (Deringer et al., 2019), (Unke et al., 2021) unlocking a variety of GBs and heterogenous interfaces which were once beyond the capability with solely AIMD, which reviewed earlier, is about 400 atoms evolving over ps of time. There are several classes of MLIPs (Behler, 2021). Second generation MLIPs, which historically have been the most widely used, express the total energy as a sum of mostly local atomic energy contributions. Third generation MLIPs include long-range interactions via environment-dependent partial charges, while fourth-generation MLIPs further account for long-range charge transfer.

For example, Lee et al. use a MLIP (system size: 600–900 atoms, deployed for 5–10 ns) and aberration-corrected scanning transmission electron microscopy and spectroscopy to study the origin of anomalously high GB conductivity in perovskite SSE, Li_0.5ySr_1-0.75y□_0.25yTa_yZr_1-yO₃ (LSTZ0.75, with A-site vacancies). Its excellent transport is unlike its canonical counterpart, perovskite-type, LLTO, which has high bulk ionic conductivity, but poor GB ionic conductivity (Lee et al., 2023). It is unclear how LSTZ0.75 and LLTO, which have the same crystal structure, can exhibit drastically different GB conductivity. A Moment Tensor Potential (MTP), fitted on DFT reference energies, is used to find that Sr at the GBs tend to migrate into the bulk for both low- Σ and high- Σ GB (Figure 2c), leaving behind Sr vacancies and promoting Li enrichment at the GBs. This observation is corroborated with dark field vibrational electron energy loss spectroscopy which finds equal concentration of Li in the bulk and at the GB. In summary, in LSTZ0.75, Li migrates as easily along the GB as through the grains due to the availability of Sr vacancies that increase promote Li transport. It is emphasized by Lee et al. that future design of high-conductivity polycrystalline SSE should purposefully include vacancy and defect engineering of the bulk SSE, as this strategy is linked to the creation of favorable GB environments for facile Li migration. This study is an example of a joint experimental-computational effort which can lead to mechanistic understandings and new materials design directions for ASSBs.

While the training error of the MLIP on GB structures is about 3 meV/atom, indicating some deviation from DFT, the accuracy is comparable to other studies using MTP on ASSB GBs (Ou et al., 2024), (Jalem et al., 2023) This could be due to the lack of long-range interactions in MTP, which could be mitigated by using a third-generation MLIP. Other opportunities include extension into systems which can only be probed using MLIPs: glassy-crystalline matrices such as Li₃PS₄ (Urata et al., 2024) and amorphous-amorphous interfaces such as Li₃B₁₁O₁₈ coatings on Li₃PS₄ SSE (Wang et al., 2023). The study of MLIP in SSE interfaces will be particularly critical in composite polymer-ceramic SSE, which is one of the most promising ASSB technologies. The reason is that composite SSE combine the advantages of high mechanical strength and interfacial contact of polymer electrolytes, with the electrochemical stability, thermal stability, ionic conductivity, and resistance to dendrite formation, of ceramic SSE (Yu and Manthiram, 2021). However, new strategies are needed to promote low-resistance, highly efficient ion-conduction networks at the composite SSE | anode and composite SSE | cathode interfaces. With this target in mind, combined atomistic and coarse-grained MLIPs may help with the design of polymer-ceramic composites.

The greatest impedance occurs not at the anode | SSE interface, but at the cathode | SSE interface, suggesting that the cathode | SSE interface could be a principal hurdle in developing ASSBs (Luntz et al., 2015). However, a main challenge is that composite cathodes, made of the cathode active material carbon binder, consist of randomly-oriented crystallographic interfaces (Figure 1b). The plethora of such interfaces implies that degradation can initiate from just as many locations.

It is unclear how to connect this complex morphology and associated surface facets to observable metrics, such as capacity fade, unless measures are taken during synthesis to intentionally generate well-defined interfaces (Zahiri et al., 2021). Recent experimental progress in this area has made this possible. Zahiri et al. performed controlled electrodeposition of LiCoO₂ with three halide SSE and one sulfide SSE. They find a strong dependence of Li₃YCl₆ instability on LiCoO₂ orientation and examine this using DFT. Their calculations reveal that the (003) LiCoO₂ surface has a weak and slightly repulsive interaction with the (100) Li₃YCl₆ surface (0.15 eV/nm²) compared to the (110) LiCoO₂ surface with (100) Li₃YCl₆ surface (4.40 eV/nm²). Since it takes less energy to break bonds at the weak interface, overpotentials will further accelerate the decomposition. In contrast, many more bonds, such as Co-Cl, Li-O, and Y-O, form at the (110) LiCoO₂ | Li₃YCl₆ interface, resisting decomposition. These insights guide rational fabrication of dense, additive-free cathodes with controlled electrochemical and chemical reactions and minimal capacity fade. This collaboration involving controlled experiments and electronic structure calculations is another example of validation efforts that enable rational, expedited optimization of ASSB.

In addition to thermodynamic comparisons, reaction dynamics at cathode | SSE interfaces have also been recently pursued. Golov et al. use AIMD to study reactivity at the (001) cathode | (001) Li₃YCl₅Br (LYCB) and (110) Li metal anode | (110) LYCB interfaces (Golov et al., 2024). Two cathode materials, LiNi_0.6Mn_0.2Co_0.2O₂ (NMC622) and LiNi_0.8Mn_0.1Co_0.1O₂ (NMC811), each containing up to ∼200 atoms, were found to promote Y-O, Li-Cl, and Li-Br bonds at the expense of Y-Cl and Y-Br bond breakage. For NMC622, 20 ps of AIMD does not reveal further structural rearrangement nor charge equilibration. On the other hand, for NMC811, more structural rearrangements are observed, specifically the formation of a bromide monochloride molecule at 15 ps, which is expected to form Br₂ gas, lead to Ni reduction, and Li intercalation from LYCB into NMC811. The findings, despite occurring over a short simulation time (15–20 ps), are consistent with experimental observations that Cl₂ gas evolves in NMC | Li₃MCl₆ (M = In, Y, Zr), and that Li loss from the SSE occurs via Li insertion into NMC. The authors suggest that dynamic evolution of NMC | LYCB can be probed, to an extent, using AIMD.

In one of the most exhaustive multiscale modeling efforts to date, Feng et al. consider 600 distinct, dense, polycrystalline LiCoO₂ | LLZO interfaces using atomistically-informed mesoscale modeling, MLIP, and ML analysis (Figure 3a). (Feng et al., 2024) Using MLIP MD simulations (sampled for 4–10 ns), they calculate effective diffusivity for 600 microstructures (consisting of a two-phase polycrystal) and find a bimodal distribution when the bulk diffusivities in LCO vs. LLZO are significantly different. This behavior arises from the respective percolating channels of slow and fast phases; when more slow-diffusing phases are added to remove connectivity of fast phases, diffusivity follows a unimodal distribution. It is revealed that microstructural optimization of heterointerfaces (defined as where at least three types of interfaces meet), can form conductive-enough networks even if diffusivity at other interfaces is low. Thus, future efforts could include collaborations with experiment, such as joint efforts to intentionally promote more of these conductive heterointerfaces. Other next steps could still be in computation, given that the authors mention three limitations in the findings: lack of atomistic detail on how spatiotemporal environments control Li mobility at interfaces, lack of charge transfer in the MLIP model, and artificial smoothing of interfacial structure for mesoscale simulations. In summary, the workflow is the first to develop atomistically-informed, multiscale modeling for cathode-active materials with SSE interfaces.

Computational efforts investigating reactivity at cathode | SSE interfaces tend to use semi-local DFT approximations, typically in the form of the Perdew–Burke–Ernzerhof (PBE) (Perdew et al., 1996), due to their ability to calculate accurate formation energies within practical cost. However, semi-local DFT contains remnant self-interaction errors (Perdew and Levy, 1983), (Mori-Sánchez et al., 2008) resulting in underestimation of band gaps and relatively poorer enthalpies of formation for strongly-correlated systems containing d electrons (Zhang et al., 2018).

One way to reduce the self-interaction error in the form of charge delocalization is to solve for on-site, Coulombic Hubbard corrections via the self-consistent linear response method (Zhou et al., 2004), (Cococcioni and de Gironcoli, 2005). However, the approach is not straightforward in cathode | SSE interfaces, as transition metal oxides near the surface have locally different environments than those in the bulk, requiring spatially and chemically dependent Hubbard U and V (DFT + U + V) corrections. These systems can benefit from automated determination, and one promising approach has been to develop ML models based on equivariant neural networks, using atomic occupation matrices to predict Hubbard parameters with iterative linear response calculations per atomic site (Uhrin et al., 2025). By removing the need for manual tuning, the method may significantly expedite ab initio modeling of cathode | SSE interfaces with significantly more heterogeneous environments.

At the interface, there is a natural accumulation of vacancies due to the difference in Li chemical potential, and this forms a space charge layer (SCL) (Figure 3b). At the anode side, there is, by contrast, an accumulation of Li and depletion of vacancies. However, depending on the SSE, the formation of a SCL does not always result in depletion of Li at the interface, since stable Li adsorption sites can form, leading to a “Li pileup”, inducing an even greater SCL. Such is the case for LiCoO₂ (110) | β-Li₃PS₄ (010). These blocked Li migration pathways can be mitigated by adding a thin LiNbO₃ buffer layer which provides more Li transport pathways, reducing the SCL (Haruyama et al., 2014).

In a follow-up study using particle swam optimization via the CALYPSO method (Gao et al., 2019), the more stable (104) surface of LiCoO₂ is systematically studied with β-Li₃PS₄ (010). It is revealed that extensive cation and anion mixing form cobalt sulfides and phosphorous oxides (Gao et al., 2020). Instead of Li adsorption forming a Li-rich region which is the case with (110) LiCoO₂ | β-Li₃PS₄ (010), there is now Li depletion due an increase in the migration barrier at the mixed sulfide/oxide interface. The resultant Li chemical potential is now higher at the interface than in the bulk (Figure 3c), resulting in interfacial resistance, electron transfer from the SSE, and oxidative decomposition of the SSE (e.g., S oxidation).

Full-scale consideration of the SCL will require simulations of large length scales with electron transfer, which are exclusively handled with at least third-generation MLIPs with environment-dependent partial charge information (Uhrin et al., 2025). While many state-of-the-art MLIPs are second generation meaning that they do not explicitly differentiate among charges or oxidation states (e.g., MTP, as discussed in the previous section), this assumption was recently shown to be problematic for the case of thermodynamic stability in bulk LiFePO₄: A pretrained CHGNet MLIP, which includes electronic degrees of freedom using magnetic moments, is successfully used with semi-grand canonical cluster expansion sampling to predict two-phase co-existence during lithiation of LiFe_xPO₄, due to its incorporation of electronic entropy (Deng et al., 2023). Without including electronic degrees of freedom, Li_xFePO₄ exists as a solid solution, inconsistent with experiments. From this work, it is evident that explicit differentiation between charge states may be critical for capturing thermodynamic behavior of cathode materials.

Even so, third-generation MLIPs with charge states for cathode | SSE interfaces may not be sufficient, considering that in an earlier section we reviewed that electron conduction in SSE can be facile along grain boundaries meaning that non-local charge transfer must also be a consideration. Fortunately, these new models, constituting fourth-generation MLIPs, are an area of active development (Ko et al., 2023) (Gubler et al., 2024).

In summary, electrochemical modeling of cathode | SSE interfaces remains an open challenge, owing to the un-established DFT corrections needed to study heterogeneous cathode environments and the unvetted non-local charge transfer MLIP models applied to cathode | SSE interfaces. Active work in these directions may lead to atomistically-informed understanding of how electrochemical degradation occurs, allowing for targeted design of passivating, ionically conducting cathode | SSE interfaces.

The cathode | cathode interface facilitates Li-ion transport within composite electrodes of ASSBs (He et al., 2021). While a well-designed particle with crystallographic alignment improves rate capability and cycling performance (Xu et al., 2020), large-scale mitigation of failure is still limited. For example, Yan et al. use STEM-HAADF to reveal that cracks nucleate from the NMC333 grain interior during cycling at high voltage (Yan et al., 2017). They find that the edge dislocation core within primary particles acts as the nucleation site for crack initiation (Figure 4a). The process is diffusion-limited, and dislocations are surmised to propagate along (003) planes, forming intragranular cracks. Indeed, crack formation along (003) is consistent with other in situ studies (Zhang et al., 2017). These findings indicate coatings are not sufficient to prevent cathode degradation at high voltage, motivating studies that focus on alleviating internal cathode strain and homogenizing Li distribution.

Achieving uniform Li concentration at increasing rates is challenging for polycrystalline cathodes as their heterogeneous structure lends to heterogeneous transport. This effect can be directly quantified with computational modeling. He et al. use AIMD to investigate Li-ion migration among different GB models and statistically analyze Li migration energy profiles in the bulk, across GBs, and along GB planes in LiNi_0.5Mn_0.3Co_0.2O₂ (He et al., 2021). They find that the coherent Σ2 GB shows the lowest boundary energy (0.097 J/m²) and promotes Li diffusivity 1-2 orders of magnitude higher than that in the bulk. However, grain boundary energy alone is not sufficient for predicting Li transport, as the Σ3 GB exhibits a modestly higher GB energy (0.286 J/m²) but reduced diffusivity by 7 orders of magnitude across the GB. Disordered Σ5 and Σ9 GBs, though energetically less favorable, promotes Li diffusivity by up to 10 times compared to the bulk.

Targeted modeling efforts can also appraise strategies to prevent intergranular cracking. Hu et al. use DFT calculations and construct a straightforward dataset for the coherent Σ3 [100] (012) GB of LiNiO₂ (Hu et al., 2024). They iterate through 64 metallic and semi-metallic dopants at four different well-defined locations: Li sites on the GB, Ni sites on the GB, neighboring Li sites near the GB, and neighboring Ni sites near the GB. Based on GB segregation energies and strength calculations, only group IV elements, such as Fe and Co, have beneficial roles as substitutes on Li sites. On the other hand, there are more effective dopant candidates for Ni GBs, suggesting that numerous GB engineering efforts (i.e., doping with Mg, Al, Si, Ti, Cr, Mn, Fe, Cu, Zn, Hf, or Ce) would be effective for Ni-rich cathodes as they would preferentially segregate and strengthen Ni GBs. The results align well with experimental observations, motivating systematic investigation across more complex systems.

The dynamic roles of dopants are already being investigated with a second-generation MLIP. Jiao et al. systematically assess the benefit of Al doping in LCO using experiments, characterization, and modeling (Jiao et al., 2025). Undoped LCO exhibits near perfect crystallization whereas doping increases chemical and structural heterogeneity. The extent that the structural heterogeneity is a problem is evaluated using a second generation MLIP. Simulations on thermodynamically favorable twin boundaries reveal that Li-ion diffusion is impeded at a LCO twin boundary compared to in the bulk (Figure 4b), resulting in loss of rate performance. Additionally, density of states DFT calculations on representative local configurations from MLIP simulations indicate that oxygen atoms with Li-rich environments are likely to be oxidized, as their electronic states lie close to the Fermi level. The iteration of using MLIP to inform configurations for electronic structure calculations is typical for closing the loop on MLIP simulation predictions.

It is worth mentioning again that Feng et al. also investigate Li transport in LCO bulk and LCO GBs within composite cathodes using MLIP (Feng et al., 2024). Their study reveals that the LCO bulk acts as a transport bottleneck, while LCO GBs can serve as secondary conduction pathways when LLZO connectivity is poor. In well-connected LLZO networks, Li-ion flux remains concentrated in LLZO. However, when LLZO pathways are disrupted, Li must traverse through the LCO bulk and GBs, resulting in reduced ionic conductivity. Their findings suggest that optimizing LCO GB properties through doping, grain size engineering, and interface modifications may enhance diffusion and improve rate and operational stability.

Some progress has been made in atomistic modeling of cathodes, mainly via DFT studies, AIMD, and MLIP. However, cathode-cathode interfaces remain understudied compared to other battery components. Thus, future directions simply include continued use of all methods to increase the scope of our understanding of cathode defects and grain boundaries during electrochemical cycling.

We refer the reader elsewhere for exhaustive Reviews on Li metal anode | SSE theory and computation (Krauskopf et al., 2020), (Sanchez and Dasgupta, 2024), (Seymour et al., 2023). In this section, we instead focus on recent works applying computational modeling to study anode-free | SSE interfaces as previous studies suggest that only an anode-free Li-metal SSB can be competitive with the energy density and cycle stability of Li-ion batteries (Heubner et al., 2021).

The anatomy of an anode-free ASSB is shown in Figure 1a. Upon charge, Li from the reservoir of the cathode active material diffuses through the thin SSE and is plated directly onto the current collector (CC). The uniformity of this plating is dependent on binding energy of Li atoms to the CC, nucleation overpotential, wettability, electric field, and morphology of the existing surface (Petersen et al., 2024).

In 2019, Pande and Viswanathan use computational screening to search for CCs which can stably nucleate uniform Li films upon charge (Pande and Viswanathan, 2019). Their thermodynamic calculations and first order approximations show that the use of Li alloys (e.g., Li-Zn, Li-Al, Li-B, Li-Cd, Li-Ag) may yield cells with better rate capability due to low nucleation overpotentials and higher specific energies.

Other thermodynamic evaluations include a combined DFT-experimental comparison of Li deposition onto Ag vs. Cu (Shin and Manthiram, 2022). Here, Shin and Manthiram find preferential interactions of Li with Ag (110) and Ag (111) surfaces, compared to Cu (111). This interaction suggests preferential alloying of Li with Ag forms Ag Li₂₀, enhancing Li reversibility and Coulombic efficiency. In another joint experimental-DFT/AIMD study with an anode-free electrode made of branched polyethyleneimine, Ag nanoparticles, and LiNO₃, deposited onto a Cu CC, Cho et al. use AIMD (size: 20 × 20 × 20 ų) to find that the Li₂O/Ag interfaces, which do not alloy as strongly compared to Li/Ag interfaces, can facilitate reversible alloying, uniform Li deposition, and increase capacity retention. However, the stable and facile formation of Li/Ag alloys are still necessary to act as buffers to prevent Li dendrite formation (Cho et al., 2022). The study demonstrates that AIMD can identify complex chemical interplay at mixed metallic/oxide interfaces in anode-free technology.

These studies of anode-free ASSB introduce an additional structure between the SSE and CC: a thin coating or intermediate layer, whether by construction (Li-alloys (Pande and Viswanathan, 2019)), or through chemical reaction (LiNO₃ decomposition (Cho et al., 2022)). The addition of this third layer has been modeled with multiscale approaches. Recently, Xie et al. combine DFT and continuum modeling to investigate the role of a thin Ag/C composite buffer layer (BL) in SSE | BL | CC (Xie et al., 2024). The computed first-principles phase diagram, including electronic energy, configurational entropy, and phonon free energy, shows remarkable solubility of Li in Ag, creating stable alloys up to AgLi₂₅. At the end of lithiation, Ag nanoparticles precipitate out of the BL onto the CC due to lower interfacial resistance, homogenizing the current. The integration of DFT with continuum modeling connects the thermodynamic underpinnings with interlayer structure to explain the observed Ag extrusion onto the CC. More efforts along this direction of “integrated atomistic to continuum mechanics” could help resolve the multiscale evolution of anode-free, buffer-layer-integrated current collectors in ASSB.

Additionally, as is the case for the cathode | SSE interface, including non-local charge transfer via MLIPs will also be critical at anode-free interfaces, as metallic Li deposition onto metallic buffer layers or current collectors means the electron is delocalized with respect to the deposited Li. For this reason, the modeling of metallic anode | SSE interfaces will also benefit from advances in fourth-generation MLIPs.

The need for more challenging datasets to include non-equilibrium solid materials has been answered by a few teams: Zheng et al. present the ab initio amorphous materials database, consisting of a ‘5000 K amorphous database’ with 5,120 compounds and with a similar coverage, chemically, with the Materials Project. Note that 69% of the compounds contain Li, making the dataset particularly suitable for ASSBs (Zheng et al., 2024).

The Open Materials 2024 (OMat24) database from Meta, consisting of 118 million single-point energy DFT PBE/PBE + U calculations on non-equilibrium inorganic bulk materials, is the largest database to date. Barroso-Luque et al. show that pre-training an Equiformer V2 Model on OMat24, with additional fine-tuning on MPtrj and subsets of the Alexandria database, yields state-of-the-art results over the top-performing U-MLIPs (ORB MPtrj, SevenNet, MACE) (Barroso-Luque et al., 2024). Overall, the availability and amount of training data appear to develop more transferable models, increasing the urgency and importance of developing more robust, open datasets.

As introduced earlier, it is possible to grow controlled morphology of thick cathode crystals using electrodeposition from a molten salt bath (at 200–300 ° C) (Zahiri et al., 2021) or epitaxial growth on a SrTiO₃ single crystal substrate (Lee et al., 2024). These capabilities can bridge the knowledge gap between crystal orientation and interfacial reactivities. For instance, it is now understood that interphase formation at the cathode | SSE interface is important for thermal stability, but that not all cathode surfaces lead to interphase growth: The (104) NMC | LLTO interface forms passive interlayers and inhibits future degradation whereas (003) NMC does not (Lee et al., 2024). As a result, (003) surfaces lead to side-reactions at elevated temperatures and extensive degradation. Future efforts bridging experiment and computation may enable targeted design of “electrochemically resilient” interfaces.

Collaborations can also provide synthesis and processing recommendations. Howard et al. carried out a joint computational-experimental study on LiMn₂O₄-LLTO cathode | SSE interface and find that cation exchange is enthalpically favorable in (111) LiMn₂O₄ | (111) LLTO such that the interface becomes a trap for anti-site defects (Howard et al., 2022). Experiments using pulsed laser deposition corroborate this finding, and observe that above a critical temperature, the interface becomes blurred. Efforts to design ASSB with sharp interfaces should consider a maximal synthesis temperature such that interdiffusion is kinetically limited. Thus, these experimental-computational studies can expedite rational understanding of ASSBs that behave “as designed”.

Experiments are typically done at constant potential, meaning constant electrochemical potential of electrons. However, typical ASSB calculations enforce constant Li chemical potential. The authors believe this is partly because grand canonical ensemble DFT, useful for enforcing constant potential, is still an area of active development. For example, recently Melander et al. describe that by enforcing a local electrode inner potential, instead of a global electrochemical potential (proportional to the Fermi level of a given phase) which is the standard method for grand canonical DFT, they can model outer-sphere reactions that occur in weakly interacting solute-electrode systems (Melander et al., 2024). Outer-sphere reactions are completely missed by the global Fermi level model. While these studies are for aqueous solid | liquid systems, they remain to be tested on ASSB interfaces. In summary, applying grand canonical DFT during AIMD studies of ASSB interfaces, and eventually to new generations of MLIPs, could also increase future collaborations with experiments since the effect of overpotentials can then directly be evaluated.

Entirely computationally-driven multiscale simulations are also being explored, and their success promises tighter integration with experiments. For example, as reviewed earlier, the Quantum Simulations Group at Livermore National Laboratory has studied chemo-mechanical effects of the garnet SSE Li₇La₃Zr₂O₁₂ | LiCoO₂ interface, guiding future processing conditions (Feng et al., 2024), (Wan et al., 2024). Validation with in operando measurements may enable construction of physically accurate, digital twin models capable of predicting and understanding battery performance in silico.