Cross-Method Benchmark of Static Surrogates and Dynamic MD Simulations

To what extent do static neural network surrogates misrank room-temperature ionic conductivity compared to machine-learning-accelerated molecular dynamics on identical candidate electrolyte datasets?

1. Introduction

The computational discovery of solid-state battery materials is rooted in a historical progression from thermodynamic frameworks to advanced machine learning techniques. Before computational methods became dominant, the field's intellectual basis emerged through the establishment of early thermodynamic and structural frameworks. Because researchers needed a formal vocabulary to describe how particles behaved within ordered phases, the field initially focused on characterizing thermodynamic stability, bond mechanisms, and rotational diffusion motion [1, 2]. This conceptual development allowed the community to classify solid ionic conductors and solid electrolytes as distinct, analyzable systems by 1982 [3]. This early period defined the macroscopic and physical chemistry categories that later computational and high-throughput methods would attempt to model and optimize.

In 1983, a methodological turning point occurred when researchers introduced first-principles calculations to derive interatomic potentials for ionic materials [4]. This allowed theoretical modeling of atomic interactions to supplement macroscopic observation, shifting the community's attention toward the practical engineering of solid state primary and secondary batteries [5]. Concurrently, new measurement techniques for electronic and ionic conductivity in high band gap materials [6] sparked a surge in empirical characterization. Throughout the 1990s and early 2000s, researchers extensively tested polymer variants [7, 8, 9] and stabilized zirconia electrolytes [10, 11]. The consequence of these parallel theoretical and experimental tracks was an iterative pipeline where physical synthesis served as the primary engine for maximizing ionic conductivity in solid-state devices.

Beginning around 2004, the unit of analysis shifted from evaluating bulk ionic conductivity to correlating precise crystal structures with specific electrochemical properties [12]. This change was caused by the growing demand for highly specific cathode and anode materials in rechargeable systems, prompting detailed modeling of electronic structure and ionic diffusion in novel layered components [13, 14]. To physically validate these increasingly complex structural requirements, researchers were forced to develop advanced preparation pipelines, moving toward advanced solid-state reaction methods and attrition milling [15]. These sophisticated techniques unlocked the ability to produce submicron and nanoscale ceramics with optimized activation energies, ensuring that synthesized physical materials could catch up to stringent theoretical conductivity targets [15].

The methodology for discovering stable candidate materials experienced a major turning point in 2020, when researchers introduced high-throughput computational screening [16]. Because manual experimental synthesis was too slow to explore the vast space of possible material compounds, researchers replaced one-by-one synthesis with atomistic simulations to systematically screen large structural repositories [16]. To execute these searches, the field came to rely on crystal structure prediction (CSP) to find energetically favorable atomic arrangements [17]. However, exploring the multidimensional potential energy surface traditionally required highly expensive density functional theory (DFT) calculations [17]. To bypass this computational bottleneck, the field recently adopted machine-learning interatomic potentials to rapidly sample and relax crystal structures [17]. This transition from manual physical synthesis to high-throughput computational screening, accelerated by machine learning models, defines the field's current momentum.

2. Theoretical Background

Building upon this historical progression, several core concepts and operational workflows define the theoretical background of the field.

Research in this field operates as a continuous loop between massive computational scree

…(truncated)

3. Statement of the Problem

Despite these methodological advancements, a fundamental problem remains regarding the predictive capabilities of surrogate models. The core question is whether machine learning surrogates can predict lithium-ion conductivity accurately without explicit ab initio molecular dynamics. This remains unresolved because graph neural networks struggle to model dynamic lattice vibrations, dynamic disorder, and temperature-dependent ion hopping pathways that dictate ionic transport. Unresolved evidence characterizes the current literature, as studies independently demonstrate high predictive accuracy for static ML descriptors and detailed mechanistic fidelity for dynamic molecular dynamics, but direct benchmark comparisons between static surrogates and dynamic simulations on identical candidate sets are lacking. Additionally, evidence is constrained by narrow chemical domains in dynamic studies, reliance on idealized crystal structures that omit grain boundary or processing effects, and surrogate features that rely on pre-computed neural network energy embeddings.

Static Representation Advocates: Static representation advocates argue that machine learning models leveraging compositional, structural, and neural network energy descriptors can accurately rank and predict solid-state ionic conductivity without requiring explicit ab initio molecular dynamics trajectories. A 2026 study supports this by applying gradient-boosted ensemble models to static structural and neural potential features across over six hundred experimental solid-state electrolytes, achieving high predictive accuracy, alongside a 2026 publication exploring machine learning predictions of ionic transport.

Dynamic AIMD Necessity Advocates: Dynamic simulation advocates argue that explicit molecular dynamics—utilizing ab initio methods or machine learning interatomic potentials—are necessary to capture complex transport behaviors such as strain effects, defect interactions, dynamic coordination changes, and amorphization. Multiple 2025 and 2026 studies support this approach, employing AIMD and machine learning interatomic potentials in molecular dynamics simulations to resolve lithium diffusion mechanisms across antiperovskites, thiophosphates, amorphized solid electrolytes, and polymer matrix systems.

4. Review of Prior Research

A detailed review of prior research contextualizes the competing positions surrounding this problem.

4.1 What is Widely Reported

Predicting ionic conductivity for solid-state electrolytes at room temperature remains a significant computational challenge [18]. Explicit ab initio molecular dynamics simulations require immensely long trajectories—typically exceeding 50 ns—to reliably capture ionic diffusion, rendering them computationally infeasible for high-throughput screening [18]. To address these computational bottlenecks, researchers utilize neural network interatomic potentials trained on density functional theory (DFT) reference data [19]. The literature establishes that machine learning potentials can reproduce DFT-level force accuracy at a substantially reduced computational cost, allowing for scalable, large-scale atomic simulations under applied electric fields [19].

4.2 What the Competing Positions Rest On

Static Representation Advocates: Advocates for static representations argue that machine learning models can accurately predict and rank solid-state ionic conductivity by utilizing compositional, structural, and pre-computed neural network energy descriptors, completely bypassing the need for explicit molecular dynamics trajectories. Research reports that a machine learning pipeline combining compositional features from the Magpie platform, structural descriptors from the OBELiX dataset, and energy features extracted from a Crystal Hamiltonian Graph Neural Network (CHGNet) achieves a test R2 of 0.775 and a mean absolute error of 0.954 log(S/cm) [20]. Applied to a benchmark dataset of 606 experimentally measured materials using XGBoost and CatBoost ensemble methods, this static approach reports performance surpassing a published random forest baseline of R2 ≈ 0.65, successfully proposing 16 novel candidate compositions consistent with the Argyrodite family [20]. Related literature also investigates the broader application of machine learning for predicting ionic conductivity in solid-state electrolytes [21].

Dynamic AIMD Necessity Advocates: Advocates for dynamic simulations maintain that explicit molecular dynamics are mandatory to capture complex transport behaviors dictated by structural disorder, dynamic coordination changes, and defect interactions. Studies establish that nonequilibrium molecular dynamics simulations accelerated by an equivariant neural network potential reveal anisotropic Li+ conduction pathways along the c-axis and mobility bottlenecks through PS4 tetrahedra in LGPS electrolytes [19]. Furthermore, simulations propose that structural modifications like doping and amorphization fundamentally alter transport mechanisms: for example, molecular dynamics using a moment tensor potential report that a 6%-Y-doped crystalline Li2ZrO3 reaches a maximum room-temperature ionic conductivity of ~2mS cm−1 [22]. Dynamic simulations also suggest the ability to capture complex synthesis trajectories, successfully reproducing the composition-dependent phase evolution of Li7P3S11, Li3PS4, and Li7PS6 from distinct Li2S precursor ratios, alongside the systematic cleavage of bridging sulfur bonds [23]. Proponents argue these dynamic phenomena—such as specific coordination shifts induced by amorphization—cannot be derived from static structural descriptors alone [24].

4.3 Foundational Work Behind the Debate

This debate stems from the inherent limitations of conventional computational methodologies used to discover materials offering higher energy density, enhanced safety, and reduced flammability [18]. Because ab initio simulations of room-temperature solid-state electrolytes demand computationally infeasible trajectory lengths, researchers historically relied on a workaround: performing molecular dynamics simulations at highly elevated temperatures and extrapolating the resulting conductivity metrics down to room temperature [18]. However, this extrapolation strategy often fails due to unpredicted phase transitions or nonphysical alterations of the materials at high temperatures [18]. The failure of elevated-temperature extrapolation created a methodological divergence in the field. One faction seeks to abandon dynamic trajectories entirely in favor of static structural machine learning models trained on large databases [20], while the opposing faction seeks to accelerate the dynamic simulations themselves using machine learning interatomic potentials to make room-temperature modeling viable [19, 23, 24].

4.4 Adjacent Debates and Neighbouring Evidence

By analogy, evidence from other domains modeling complex material behaviors demonstrates the broad applicability and transferability of machine learning interatomic potentials for dynamic phenomena. In the study of nanoporous materials, deep neural network potentials trained on a vast density functional theory dataset (350,000 unique calculations across 219 topologies) report impressive agreement for predicting structural properties, energy-volume trends, and phonon density of states in pure silica zeolites [25]. In a different setting exploring material failure, artificial neural network potentials report success in describing the individual bond-breaking events required for modeling mechanics and fracture dynamics in covalent crystals like graphene, utilizing density-functional tight-binding methods for extensive data generation [26].

Similarly, investigations into phase stability suggest that machine learning potentials can robustly model extreme dynamic transitions. Studies of calcium oxide report that neural network potentials successfully capture the material's melting behavior, high-pressure melting curves, and a ~29% volume increase at the melting temperature [27]. Furthermore, research on liquid alkali metal alloys (RbxNa1-x) establishes that machine learning potentials based on ab initio molecular dynamics can resolve the delicate balance between energy and entropy contributions, uncovering asymmetric diffusion dynamics and specific atomic clustering behaviors [28]. These adjacent findings imply that machine learning interatomic potentials are highly capable of resolving complex, time-evolving structural changes across diverse physical chemistry domains, reinforcing the dynamic simulation advocates' position that explicit dynamic modeling is both scalable and necessary.

4.5 Contradictory and Inconsistent Findings

The literature reveals stark inconsistencies regarding how amorphization and structural disorder impact ionic conductivity, contradicting the assumption that universally applicable static rules govern ion transport. Molecular dynamics simulations utilizing machine learning interatomic potentials report that amorphization increases ionic conductivity for Li3YCl6, decreases conductivity for Li0.33La0.56TiO3, and leaves conductivity relatively unchanged for Li7P3S11 [24]. Coordination analysis at 300K suggests these divergent responses occur because amorphization induces distinct changes in Li-anion coordination environments, demonstrating that the effect of structural disorder on Li-ion diffusion behavior is not universal [24].

Further inconsistencies appear when assessing defect interactions. While amorphization reportedly enhances the conductivity of undoped Li2ZrO3, this enhancement is actively attenuated when the material is doped [22]. Additionally, amorphous Y-doped Li2ZrO3 exhibits spatially fragmented, trapping-dominated lithium transport, in direct contrast to its crystalline counterpart [22]. These findings indicate that simple structural descriptors may fail to predict conductivity accurately across different anion chemistries because specific, highly localized coordination motifs ultimately dictate the diffusion behavior [24].

4.6 Limitations of the Existing Evidence

The retrieved evidence base exhibits distinct limitations that prevent a conclusive resolution. The most critical limitation is the absence of direct benchmark comparisons between static surrogate models and dynamic simulations applied to identical sets of candidates. While static descriptors are evaluated on broad datasets encompassing hundreds of materials [20], explicit dynamic simulations rely on models trained and validated within narrow chemical spaces, such as specific thiophosphates [23, 24], Li2ZrO3 cathode coatings [22], or LGPS structures [19]. Additionally, models predicting synthesis trajectories report that limited simulation timescales currently restrict full crystallization, limiting their outputs to glassy intermediates rather than fully resolved crystal phases [23]. Finally, dynamic simulations heavily rely on idealized crystalline structures or specific defect formulations, omitting the influence of grain boundaries, secondary phase formation, and macroscopic solid-state battery processing realities.

4.7 What Remains Unresolved

The literature genuinely cannot settle whether the neural network energy features embedded in static representations—such as those extracted from CHGNet [20]—inherently encode sufficient information about dynamic lattice v

…(truncated)

5. Research Gap and Purpose of the Study

Addressing what remains unresolved, this study identifies a specific research gap and proposes a targeted opportunity. The literature lacks direct benchmark comparisons evaluating static surrogate models against machine learning interatomic potential molecular dynamics on identical, standardized sets of candidate solid electrolytes across diverse chemical spaces. Researchers can construct a standardized, multi-class electrolyte benchmark dataset (spanning thiophosphates, oxides, and halides) to systematically measure predictive accuracy and candidate ranking discrepancies between static neural potential surrogates and long-timescale molecular dynamics. Unresolved evidence indicates that static surrogates [20] and dynamic simulations [19, 24] have been evaluated independently on disconnected chemical domains without direct head-to-head performance benchmarks on identical candidate sets. This gap is critical; if the assumption that static surrogate models reliably predict candidate performance proves false, materials discovery leads will misallocate laboratory resources toward candidate solid electrolytes that fail to demonstrate expected transport properties under dynamic conditions. Experimental chemists and battery development engineers relying on static ranking pipelines would consequently invest synthesis efforts into candidates whose activation barriers or diffusion bottlenecks are obscured by non-dynamic representations. Conversely, if static representations are falsely assumed to be insufficient when they are actually reliable, computational researchers will incur unnecessary high-performance computing expenses by running prolonged molecular dynamics trajectories across large candidate spaces.

6. Research Questions and Hypotheses

To address this gap, the study is guided by the following primary research questions and hypotheses.

Primary Question: To what extent do static neural network surrogates misrank room-temperature ionic conductivity compared to machine-learning-accelerated molecular dynamics on identical candidate electrolyte datasets?

Sub-question 1: Are ranking discrepancies systematically clustered within specific solid-state electrolyte chemical classes (e.g., thiophosphates, oxides, halides)?

Unit of Analysis: Individual candidate electrolyte crystal structure.

Level of Analysis: Material level.

Hypotheses

(Note: These hypotheses are mutually exclusive. H1 serves as the rival to H2.)

7. Method

The following methodology outlines the research design constructed to test these hypotheses.

7.1 Theoretical Framework

This design sits on the theoretical divergence regarding atomic transport in solid-state electrolytes [29]. Historically, models extrapolated room-temperature performance from elevated-temperature simulations, which frequently failed due to unforeseen phase transitions. This failure generated two competing theoretical frameworks for computational materials discovery. The first framework posits that ionic conductivity can be effectively represented through static structural, compositional, and energy-based descriptors without requiring explicit time integration [30, 31]. The second framework argues that complex ionic transport is fundamentally dynamic; properties such as anisotropic conduction pathways, strain effects, defect interactions, and amorphization-induced coordination shifts fundamentally dictate diffusion bottlenecks [32, 33, 34, 35]. If the dynamic framework holds true for materials discovery, static surrogates will systematically fail to capture kinetic barriers [32, 33, 34], necessitating ML-accelerated molecular dynamics to evaluate ionic diffusion [36, 37]. However, it must be explicitly stated that ML-MD is a computational proxy with its own inherent potential fitting errors and constrained chemical domains, preventing it from serving as an absolute physical ground truth for ionic conductivity.

7.2 Critical Assumptions

Reality Checks

Open Methodological Questions

7.3 Variables

7.4 Operational Definitions

7.5 Data

1. Assemble and standardize the dataset of .cif files across multiple classes.

2. Attempt to identify and isolate structures to mitigate data leakage between candidates and the static model's training distribution.

3. Generate and mathematically lock all predictions from the static neural network surrogate pipeline.

4. Execute the ML-MD simulations on the identical dataset, identifying non-convergent trajectories.

5. Calculate dynamic conductivity estimates and compile the final comparative dataset using dropwise deletion for materials failing dynamic simulation.

7.6 Methodology

Design type: Correlational, paired-prediction benchmark study.

Procedure:

1. Preparation: Standardize the subset of crystal structures across the designated chemical families.

2. Static Evaluation: Pass all candidate structures through the static surrogate model. Extract and record the predicted Log(S/cm).

3. Contamination Control (Locking): The dataset of static predictions must be pre-registered and cryptographically locked (or securely archived by an independent researcher) prior to initiating any dynamic simulations. This design-level separation ensures zero post-hoc tuning of the static surrogate based on dynamic results.

4. Dynamic Evaluation: Run the identical structures through explicit machine-learning-accelerated molecular dynamics (using the predetermined ML potential). Ensure identical thermodynamic parameters (temperature, pressure) for all runs. Extract the final calculated ionic conductivity.

5. Alignment: Rank the materials independently based on both output distributions. Execute the planned associative analyses on the differences.

Identification Strategy: The study isolates the methodological discrepancy by deploying perfectly paired evaluations (static vs. dynamic) on universally identical input datasets (identical crystal structures).

What it can identify: The correlational relationship between static surrogate predictions and dynamic ML-MD predictions; the specific structural families where systematic misranking occurs; the magnitude of the prediction discrepancy across the solid-state electrolyte classes.

How it avoids the unresolved flaw: By forcing a direct, 1:1 cross-method comparison on an identical candidate set, this design bypasses the current literature's reliance on disparate evaluations across non-overlapping chemical domains.

7.7 Analysis Plan

Layer 1: Readable Form

Full specification

Terms:

Supplementary Ranking Analysis:

To directly test the primary hypotheses regarding misranking, a Spearman rank correlation coefficient ($\rho$) will be computed between the static and dynamic rank vectors. Differences in absolute rank position will be assessed via paired non-parametric difference tests (e.g., Wilcoxon signed-rank test).

Robustness Checks & Assumptions:

8. Expected Contributions

By executing this methodology, the study anticipates several significant theoretical and practical contributions. If H1 (Significant misrank) is supported: It provides empirical justification for the dynamic simulation advocates, signaling to experimental chemists and batt

…(truncated)

9. Limitations

Despite its rigor, the proposed design is subject to specific methodological limitations. The study cannot identify true experimental, laboratory-synthesized ionic conductivity, nor can it identify the physical causal mechanisms behind the algorithmic discrepancies without subsequent mechanistic probing.

10. Feasibility

The successful execution of this research depends on several logistical and computational factors.

Overall Feasibility: The study is methodologically feasible provided the necessary open-source pre-trained static architectures and requisite institutional computing power for the machine-learning molecular dynamics baseline can be verified prior to execution.

References