EDIS: Ion Intercalation/Deintercalation Simulator

• EDIS is a modular simulation platform designed to study atomistic ion intercalation/deintercalation in battery electrodes with machine-learning-enhanced energetics.
• The platform employs sequential insertion/extraction workflows, structure optimization, and molecular dynamics to capture SOC-dependent evolution, voltage profiles, and transport behavior.
• EDIS integrates tools like ASE, Pymatgen, LAMMPS, and deep learning models to analyze concentration redistribution, defect formation, and phase transitions during cycling.

Electrode Dynamic Ion Intercalation/Deintercalation Simulator (EDIS) is a modular simulation and analysis platform for atomistic studies of ion insertion and extraction in battery electrode materials, developed primarily for lithium-ion battery electrodes and designed to bridge the gap between highly accurate but expensive quantum-mechanical calculations and efficient but often insufficiently accurate classical simulations. Its central aim is to track dynamic ion intercalation/deintercalation, concentration redistribution, structural relaxation, phase transitions, defect formation, and transport-property evolution across state of charge (SOC), using machine-learning interatomic potentials together with workflow automation and post-processing modules tailored to electrode materials (Wang et al., 14 Aug 2025).

1. Scientific scope and target phenomena

EDIS is motivated by the observation that electrode materials do not act as static hosts during charging and discharging. Instead, they undergo continuous ion intercalation/deintercalation, concentration redistribution, structural relaxation, phase transitions, defect formation, and transport-property changes, and these microscopic processes control macroscopic battery metrics such as capacity, voltage profile, rate capability, stability, and degradation (Wang et al., 14 Aug 2025).

The platform is framed around unresolved microscopic phenomena in both cathodes and anodes. For layered cathodes such as $\mathrm{LiNiO}_2$, deep delithiation can trigger cation disordering, especially Ni migration into Li layers, and lattice oxygen loss associated with highly oxidized Ni species and structural instability. For graphite anodes, repeated lithiation/delithiation leads to staging phenomena, interlayer-spacing changes, concentration-dependent diffusion, and possible defect or stacking-fault accumulation. EDIS is therefore designed to investigate not only equilibrium structures but also the dynamic coupling among Li concentration, Li spatial distribution, local occupancy, lattice-parameter evolution, structural defects, transport behavior, and cycling-induced irreversibility (Wang et al., 14 Aug 2025).

Its detailed demonstrations focus on lithium-ion systems, especially the $\mathrm{LiNiO}_2$ cathode and graphite anode. The paper states that, with simple parameter adjustment, the platform can be extended to sodium-ion batteries and related systems, but it does not provide a sodium-specific case study or sodium-specific equations (Wang et al., 14 Aug 2025).

2. Architecture, implementation, and module structure

EDIS is implemented primarily in Python and functions as an automation and analysis layer around established atomistic software ecosystems rather than as a monolithic standalone molecular-dynamics engine. ASE is used for structure I/O, cell operations, and optimization; Pymatgen for structure parsing and species handling; LAMMPS for molecular dynamics; dpdata and DeepPot for machine-learning model prediction; Plotly for interactive 3D insertion-energy maps; Matplotlib for publication-style plots; Pandas and NumPy for data processing; and OVITO for trajectory and structural visualization. In the optimization module, CHGNet is explicitly used as a pretrained universal neural-network potential, while insertion and extraction modules use DeepPot prediction through dpdata.System(...).predict(dp_model) (Wang et al., 14 Aug 2025).

The platform is organized into eleven modules.

Category	Modules	Role
Core	insert.py, extract.py	Site screening and sequential Li insertion/extraction
Auxiliary	opt.py, get_msd.py, poscar_to_lammps_input.py, lammps_dump_to_poscar.py	Relaxation, MSD extraction, and format conversion
Property analysis	count_event.py, concentration_vs_soc.py, msd_vs_soc.py, cell.py, voltage.py	SOC-resolved event, concentration, transport, cell, and voltage analysis

The software begins from structure files, typically in VASP POSCAR format. A structure may first be optimized with opt.py, which wraps an ASE calculator and the BFGS optimizer. The fix_cell_angles utility keeps lattice angles fixed while allowing cell lengths and atomic coordinates to relax. Dynamic charge/discharge simulations are then driven by shell scripts that repeatedly call extract.py followed by opt.py for delithiation, or insert.py followed by opt.py for lithiation. If molecular dynamics is performed, trajectories are generated in LAMMPS and converted back to VASP-style snapshots with lammps_dump_to_poscar.py, while poscar_to_lammps_input.py handles format conversion in the reverse direction (Wang et al., 14 Aug 2025).

3. Simulation workflow and principal algorithms

Methodologically, EDIS combines structure optimization, site screening for Li insertion and extraction, molecular dynamics, and SOC-resolved post-processing. Its central computational premise is that a trained machine-learning potential can predict energies and forces for large numbers of candidate structures at much lower cost than direct first-principles calculations, while maintaining accuracy “approaching that of quantum mechanical methods in relevant chemical spaces” (Wang et al., 14 Aug 2025).

The workflow proceeds in a sequence. First, an initial structure is read using ASE or Pymatgen. Second, the structure can be optimized with ASE, a machine-learning calculator, and BFGS, with variable cell lengths and fixed cell angles. Third, EDIS performs automated, stepwise Li extraction or insertion. Fourth, after each event, the structure can be re-optimized and passed to the next cycle, thereby building a sequential charge/discharge pathway. Fifth, molecular dynamics via LAMMPS can be run on optimized or intermediate structures. Sixth, analysis modules process SOC-resolved structures and trajectories to compute Li layer concentrations, insertion/extraction event distributions, mean squared displacement (MSD) trends, voltage curves, and cell-parameter evolution (Wang et al., 14 Aug 2025).

The insertion algorithm uses a grid-search strategy. Unit-cell dimensions are read, grid counts along each axis are chosen so that the grid spacing is no smaller than a user-defined value such as $0.4~\text{\AA}$, and candidate points too close to existing atoms are removed using a minimum-distance criterion such as $1.5~\text{\AA}$. For each remaining point, a Li atom is inserted, the candidate structure is written, and the total energy is predicted with the machine-learning model. The lowest-energy point is selected as the preferred insertion site, and all points and energies can be visualized as an interactive 3D heatmap. The geometric filter is expressed operationally as

$$\min_j \left| \mathbf{r}_{\mathrm{grid}}-\mathbf{r}_j \right| > 1.5~\text{\AA}.$$

The extraction algorithm uses what the paper calls a “queue-bubbling method”: all possible one-Li-removed structures are generated, their total energies are predicted, and the lowest-energy candidate is ranked to the top and selected as the preferred extraction event (Wang et al., 14 Aug 2025).

These operations are repeated in shell-script loops to mimic continuous lithiation or delithiation. The resulting pathway is a sequential low-energy pathway construction under repeated local energy screening plus relaxation. The paper is explicit that this is not kinetic Monte Carlo or transition-state theory, and it does not report explicit transition-state searches, nudged elastic band calculations, defect thermodynamics formulas, or migration-barrier calculations (Wang et al., 14 Aug 2025).

4. State variables, descriptors, and quantitative observables

The principal state variables and operating descriptors in the paper are lithium content or composition $x$ in $\mathrm{Li}_x\mathrm{NiO}_2$, SOC, Li spatial distribution across layers, total energy, unit-cell parameters $a$, $b$, and $c$, MSD, voltage, and structural snapshots during cycling. Temperature, pressure, timestep, and sampling interval are adjustable parameters when molecular dynamics is run through LAMMPS, but the example calculations do not provide a specific thermodynamic protocol matrix (Wang et al., 14 Aug 2025).

For voltage analysis, EDIS uses explicit energy normalization. If the structure contains $\mathrm{LiNiO}_2$0 atoms and $\mathrm{LiNiO}_2$1 Ni atoms, and the machine-learning model provides an energy per atom $\mathrm{LiNiO}_2$2, then

$\mathrm{LiNiO}_2$3

For two neighboring compositions $\mathrm{LiNiO}_2$4 and $\mathrm{LiNiO}_2$5 in $\mathrm{LiNiO}_2$6, with energies per formula unit $\mathrm{LiNiO}_2$7 and $\mathrm{LiNiO}_2$8, and Li-metal reference energy $\mathrm{LiNiO}_2$9, the segmental voltage is

$0.4~\text{\AA}$0

In the code, $0.4~\text{\AA}$1 eV (Wang et al., 14 Aug 2025).

Transport analysis is centered on MSD. The core quantity is

$0.4~\text{\AA}$2

and the paper states that statistical analysis of MSD from molecular dynamics provides lithium-ion diffusion coefficients and reveals migration paths. The get_msd.py and msd_vs_soc.py modules process element-resolved LAMMPS output (out.msd) and align data across SOC windows or cycles (Wang et al., 14 Aug 2025).

For layered materials, EDIS also performs layer-resolved concentration analysis using K-means clustering on fractional coordinates along the layered direction, typically the $0.4~\text{\AA}$3-axis. Host-layer positions are partitioned into layer clusters, and Li positions are counted between adjacent cluster centers to estimate layer-resolved occupancy. This procedure is central to how the software converts atomistic trajectories into physically interpretable descriptors such as per-layer Li concentration (Wang et al., 14 Aug 2025).

5. Demonstrated applications and representative findings

The principal case study is $0.4~\text{\AA}$4. EDIS demonstrates variable-cell optimization of a $0.4~\text{\AA}$5 supercell using CHGNet, then automated continuous Li extraction down to $0.4~\text{\AA}$6, followed by reverse insertion back to $0.4~\text{\AA}$7. Intermediate structures shown include $0.4~\text{\AA}$8, $0.4~\text{\AA}$9, $1.5~\text{\AA}$0, and $1.5~\text{\AA}$1. This sequentially updated pathway allows structural evolution to be tracked continuously across SOC instead of inferred from isolated snapshots (Wang et al., 14 Aug 2025).

The concentration analysis reveals a pronounced asymmetry between extraction and insertion. During extraction from $1.5~\text{\AA}$2, Li removal probabilities from different layers are approximately equal, so all layers decrease in concentration nearly uniformly. During insertion, however, the process is spatially inhomogeneous: Layer 2 and Layer 3 fill faster and saturate earlier than the others. The unit-cell analysis further shows that the $1.5~\text{\AA}$3 and $1.5~\text{\AA}$4 axes vary continuously and monotonically, whereas the $1.5~\text{\AA}$5-axis exhibits non-monotonic behavior during insertion, indicating a more complex out-of-plane response (Wang et al., 14 Aug 2025).

The voltage module produces a delithiation voltage curve spanning roughly the experimentally expected range, shown from about $1.5~\text{\AA}$6 to $1.5~\text{\AA}$7 V depending on composition, and the paper states that the calculated $1.5~\text{\AA}$8 voltage profile matches the experimentally measured voltage range closely (Wang et al., 14 Aug 2025).

Graphite is treated as the representative anode problem in the theoretical framing and in the conclusion. The paper states that EDIS was also used to simulate graphite lithiation/delithiation and obtain insight into the staging transformation mechanism (Wang et al., 14 Aug 2025). The class of graphite phenomena that the platform is intended to address includes stage-structure formation, interlayer-spacing changes, concentration-dependent diffusion, and defect or stacking-fault accumulation. A related machine-learning-potential study of graphite resolves stage transitions driven by carbon-layer sliding and reorganization, defect-regulated transport, and a kinetic asymmetry between intercalation and deintercalation, which clarifies the kind of atomistic graphite physics that an EDIS-type workflow can probe in detail (Wang et al., 8 Aug 2025).

More generally, the platform is intended to expose how preferred insertion and extraction sites, SOC-dependent total energy, voltage–composition curves, lattice-parameter evolution, layer-resolved Li occupancy, insertion/extraction event frequencies, element-resolved MSD curves, and structural trajectories are jointly related. This suggests a workflow for evaluating concentration heterogeneity, insertion/extraction asymmetry, vulnerable deep-delithiation regimes, and the impact of disorder or local reconstruction on transport and electrochemical behavior (Wang et al., 14 Aug 2025).

6. Position within the battery-modeling landscape and current limitations

EDIS occupies the atomistic end of the battery-modeling spectrum. It differs from Cahn–Hilliard reaction models that resolve phase separation inside isotropic particles (Zeng et al., 2013), from modified porous electrode theory that embeds free-energy-based multiphase thermodynamics into electrode-scale transport (Ferguson et al., 2014), and from full-cell phase-field formulations that resolve porous-electrode microstructure and concentration-front propagation at the continuum scale (L'vov et al., 2021). It also differs from interface-kinetics frameworks that modify Butler–Volmer kinetics to capture ultrahigh-rate intercalation reactions (Xiong et al., 2022). In contrast, EDIS focuses on atomistic structural evolution, site selection, defect-sensitive pathways, and transport descriptors extracted from machine-learning-potential-based structural optimization and molecular dynamics (Wang et al., 14 Aug 2025).

This positioning makes EDIS complementary to endpoint or constitutive workflows rather than redundant with them. High-throughput screening methods can predict volume change upon intercalation for large sets of candidate materials, but they do not model diffusion, time dependence, hysteresis, path dependence, or cycle-by-cycle evolution (Baumann et al., 11 Mar 2026). Equilibrium DFT analyses can decompose electrode volume change into ionic, electronic, and host contributions and thereby inform chemo-mechanical constitutive laws, but they do not provide dynamic insertion/deintercalation pathways or transport trajectories (Maréchal et al., 2024). EDIS addresses the missing atomistic pathway layer between such equilibrium screening strategies and larger-scale continuum battery models (Wang et al., 14 Aug 2025).

The present version also has clear limitations. The paper does not provide an active-learning loop, uncertainty estimation strategy, autonomous retraining workflow, loss functions, descriptor definitions, neural-network architectures for the deployed potentials, or benchmark errors such as energy RMSE and force RMSE for EDIS itself. It does not include explicit transition-state or barrier-search tools, generalized defect thermodynamics modules, detailed uncertainty-aware machine-learning-potential management, or direct coupling to larger-scale chemo-mechanical models. Its current K-means and layered-structure logic is especially suited to ordered layered materials such as $1.5~\text{\AA}$9 and graphite, and more general algorithms are said to be needed for multiphase, interfacial, or highly disordered systems. The paper also identifies future integration of big-data visualization and AI-assisted analysis as a direction for increasing the platform’s intelligence and usability (Wang et al., 14 Aug 2025).

In this sense, EDIS is best understood as an end-to-end atomistic workflow for SOC-resolved dynamic intercalation/deintercalation studies, rather than as a universal battery simulator. Its distinctive contribution is the integration of machine-learning-potential energetics, sequential insertion/extraction workflow automation, molecular-dynamics-based transport analysis, and electrode-specific post-processing in a form tailored to microscopic mechanism discovery in intercalation materials (Wang et al., 14 Aug 2025).