Why Conventional AIMD Falls Short for Solid Electrolyte Interface Prediction

Conventional ab initio molecular dynamics (AIMD) cannot simulate the solid electrolyte interface (SEI) at the time and length scales that matter for battery engineering. The SEI is a complex, heterogeneous passivation layer that forms in situ on battery electrodes — electronically insulating but ionically conducting — and its composition, structure, and formation mechanism remain poorly understood despite decades of research. As framed in the literature, the challenge is a “trade-off between accuracy and accessible time- and length scales”: DFT-level accuracy is too expensive for the system sizes and timescales needed to describe SEI growth realistically.

Patent evidence corroborates this directly. Both Hyundai Motor Company’s 2024 US filing and Kia Corporation’s corresponding Chinese filing explicitly state that AIMD suffers from high computational cost, small predicted crystal structures, and the need to simulate at artificially elevated temperatures rather than room temperature — all of which compromise the accuracy and practical utility of predictions. Classical MD without bespoke interatomic potentials is similarly limited: it cannot be applied to novel compositions where force fields do not exist.

Foundational AIMD work by Sandia National Laboratories on the initial stages of SEI formation on graphite anodes demonstrated that ethylene carbonate (EC) decomposition is kinetically controlled and highly sensitive to carbon edge terminations — results that required explicit liquid EC/graphite interface simulations. Such calculations, while physically illuminating, scale poorly and cannot be routinely applied at engineering scale. According to the U.S. Department of Energy, battery materials simulation is one of the most computationally demanding domains in applied materials science, a reality that motivates the entire AI-acceleration agenda reviewed here.

Machine Learning Potentials: Near-DFT Accuracy at Orders-of-Magnitude Lower Cost

Machine learning interatomic potentials (MLIPs) are surrogate models trained on DFT data that evaluate energies and forces several orders of magnitude faster than quantum-mechanical calculations — enabling the time and length scales needed for realistic SEI modeling. Columbia University’s comprehensive survey of these methods notes that ML potentials trained on accurate first-principles data “enable computationally efficient linear-scaling atomistic simulations with an accuracy close to the reference method,” and that ML-based property-prediction and inverse design techniques are powerful for computational search in solid-state battery materials including electrolytes, electrode-electrolyte interfaces, and SEI components.

“Energies and forces can be computed several orders of magnitude faster than DFT without loss in accuracy” — University of Göttingen, on high-dimensional neural network potentials applied to LixMn₂O₄/water interfaces.

A particularly well-validated implementation comes from UC San Diego, where moment tensor potentials (MTPs) trained on DFT energies and forces using the van der Waals optB88 functional yielded accurate lattice parameters for Li₀.₃₃La₀.₅₆TiO₃, Li₃YCl₆, and Li₇P₃S₁₁. NPT MD simulations using these MTPs correctly predicted ionic conductivities and activation energies, bridging the longstanding gap between AIMD predictions and experimental measurement — a gap attributed to the artificially high temperatures and short timescales imposed by computational constraints in conventional AIMD.

High-dimensional neural network potentials (HDNNPs) have been applied specifically to solid-liquid battery interfaces by the University of Göttingen, enabling ML-driven simulations of interfaces between water and LixMn₂O₄ — including electronic structure analysis via a companion neural network for spin prediction. This result directly demonstrates that ML potentials extend accessible time and length scales to regimes necessary for realistic interface modeling. The University of Oxford has framed the broader argument: quantum-mechanical methods “quickly reach their limits when complex electrochemical systems are to be studied — for example, when structural disorder or even fully amorphous phases are present, or when reactions take place at the interface between electrodes and electrolytes.” ML-based interatomic potentials are “many orders of magnitude faster” while affording quantum-mechanical accuracy.

From the patent side, Quantum Generative Materials LLC has filed a US pending patent for a machine-learning-driven framework that trains on molecular structures with intrinsic atomic features and atomic weights, applies bias correction to improve training, and outputs MD simulation results for target solid-state electrolyte structures. Hyundai Motor Company’s approach uses machine learning to calculate a potential specific to a simulated crystal structure and then deploys that potential within classical MD — explicitly designed to resolve the speed-accuracy trade-off of AIMD. Both represent direct productization of the MLIP paradigm, as tracked by PatSnap’s R&D intelligence platform.

SEI-Specific Simulation Strategies: Reactive MD and Kinetic MC-MD Hybrids

Beyond general MLIPs, several specialized simulation methodologies have been developed or validated specifically for the SEI, each targeting a different aspect of the accuracy-cost trade-off. The Dassault Systèmes kinetically corrected Monte Carlo-MD hybrid approach uses non-reactive force fields with graphically defined reactions of arbitrary complexity, deriving reaction probabilities from ab initio calculations. This method successfully reproduces the experimentally observed two-layered SEI structure on graphite — an inorganic inner layer and organic outer layer — developing via a near-shore aggregation mechanism. By decoupling reaction probability calculation (performed once at the ab initio level) from the dynamical evolution (handled by classical MD), the approach achieves significant cost reduction while retaining chemical realism.

Reactive force fields (ReaxFF) provide an alternative route to capturing SEI chemistry within MD. Soochow University’s review of ReaxFF studies of sulfur cathodes, various anodes, and electrolytes emphasizes the force field’s ability to describe both physical processes and chemical reactions in systems significantly larger than those accessible to quantum chemistry. Kiel University used ReaxFF to investigate SEI formation on silicon oxide surfaces under ether-based, high-concentration LiTFSI electrolyte conditions — with results verified against galvanostatic testing and XPS — demonstrating that validated reactive MD can substitute for far more expensive AIMD in mechanistic SEI studies. Research published by Nature and affiliated journals has consistently highlighted reactive force field approaches as a practical bridge between quantum accuracy and engineering-scale simulation.

Chinese institutional research has addressed artificial SEI (ASEI) modeling through two related pending CN patents from Shenzhen Power Supply Bureau. The first details a workflow starting from crystal structure data, building heterojunction structures with lattice matching, calculating lithium-ion adsorption sites, and simulating migration pathways — a complete computational pipeline for ASEI interface analysis. The companion patent focuses on performance parameter evaluation through the same heterojunction/lattice-matching framework, aiming to increase accuracy of ASEI assessment. Both patents represent an institutionalization of ML-augmented interface simulation workflows in industrial battery R&D.

Multi-Task Learning and High-Throughput Screening: Maximising Information from Cheap Simulations

A distinct but related line of work leverages AI not to accelerate individual MD runs, but to reduce the total simulation budget required for materials screening by learning from noisy or incomplete simulation data. MIT’s Department of Mechanical Engineering trained a multi-task graph neural network on a large corpus of short, unconverged MD data and a small set of long, converged reference simulations. The model achieves accurate predictions of four converged properties across a chemical space of 6,247 polymers — described as “orders of magnitude larger than previous computational studies” — by learning to compensate for the noise and non-convergence inherent in cheap, short simulations.

This strategy directly addresses the two main cost drivers in amorphous electrolyte MD: the need for multiple repeated sampling runs (due to structural disorder) and the requirement for long simulations (due to slow relaxation). The approach is particularly significant because amorphous phases are common in SEI layers, making conventional convergence criteria especially expensive to satisfy. Standards bodies including NIST have flagged amorphous material simulation as one of the most computationally demanding frontiers in battery materials science.

High-throughput screening of inorganic solid electrolytes has also been accelerated through hierarchical ion-transport algorithms, as described by the Australian Nuclear Science and Technology Organisation. The SPSE platform integrates a materials database with empirical geometric analysis and bond valence site energy methods to pre-screen candidates before invoking expensive first-principles nudged elastic band (FP-NEB) calculations, automating the preprocessing pipeline and reducing the compute-intensive steps to only the most promising candidates. This hierarchical approach mirrors the broader principle of using cheap ML or empirical filters to reduce the number of candidates requiring expensive simulation — a strategy that, according to WIPO‘s technology trend reports, is increasingly embedded in battery materials IP filings globally.

Machine learning models that predict conductivity directly from descriptors — bypassing MD altogether for initial screening — have been demonstrated across multiple studies. University College London used logistic regression on elemental feature descriptors derivable from tabulated unit cell and atomic property data to screen NASICON-type compounds. The Faraday Institution’s review synthesises common ML techniques including Gaussian process regression, neural networks, and random forests for solid-state electrolyte discovery, noting that these approaches could “accelerate the process significantly” beyond conventional trial-and-error. Columbia University’s interface modeling review establishes that ML models are “computationally much more efficient than first principles methods” and enable modeling of “larger systems and extended timescales, a necessary prerequisite for the accurate description of many interface properties.”

Key Players and Innovation Trends: From National Labs to Automotive OEMs

Analysis of assignee frequency and technical depth across the dataset reveals a clear tier structure of institutional activity in AI-accelerated SEI simulation. Automotive OEMs, commercial AI-materials startups, national laboratories, and academic groups are each contributing distinct components of the emerging technology stack.

Automotive OEMs: Hyundai Motor Company and Kia Corporation

Two filings from affiliated Korean automotive entities and a third US filing reflect a sustained industrial commitment to ML-accelerated solid electrolyte conductivity prediction, directly motivated by solid-state battery development for electric vehicles. The explicit identification of AIMD’s limitations — high cost, small system sizes, elevated temperature requirements — in these patent documents signals that OEMs are actively seeking IP protection for the solutions, not just documenting the problem.

Dassault Systèmes: Multi-Scale Simulation Frameworks

Two literature contributions from Dassault Systèmes on kinetically corrected Monte Carlo-MD SEI growth and multi-scale electrolyte transport simulations position the company as a leading provider of multi-scale simulation frameworks for battery electrolyte modeling, spanning from atomistic MC-MD SEI growth to cell-level electrochemical models. This breadth of coverage — from the nanoscale SEI to the device level — is a distinctive commercial capability.

Sandia National Laboratories: Foundational AIMD Benchmarks

Three literature references from Sandia establish the laboratory as the foundational contributor to AIMD-based SEI chemistry, providing the high-accuracy reference benchmarks against which ML-accelerated methods are validated. Sandia’s AIMD studies of EC decomposition on graphite and electronic structure modeling of electrode/electrolyte interfaces represent the “ground truth” that the ML community targets when claiming near-DFT accuracy. The PatSnap Insights database tracks Sandia’s continued output as a benchmark setter in this domain.

Commercial and Startup Activity: Quantum Generative Materials LLC and Shenzhen Power Supply Bureau

Pending patents from Quantum Generative Materials LLC and Shenzhen Power Supply Bureau signal that AI-accelerated solid electrolyte simulation is transitioning from academic research to industrial IP protection. Quantum Generative Materials represents an emerging commercial player specifically targeting AI-accelerated solid electrolyte discovery as a product offering; Shenzhen Power Supply Bureau’s two CN patents on ASEI simulation and performance evaluation reflect Chinese utility sector interest in computational ASEI design for grid-scale lithium metal batteries.

“Understanding the SEI’s precise formation mechanism, structure, composition, and evolution remains a conundrum — and data-driven modeling is the most promising pathway to resolving it at the time and length scales that matter.”

Innovation trends visible across the dataset include: a shift from single-property AIMD to multi-property ML surrogate models; increasing use of bias-correction and multi-task learning to maximise information extracted from limited expensive data; integration of ML-predicted potentials with classical MD engines rather than replacement of MD entirely; and growing patent activity from automotive OEMs and commercial AI-materials startups alongside traditional academic and national laboratory contributors. Tracking these trends systematically requires access to both patent and literature databases — a task well suited to platforms such as PatSnap‘s integrated innovation intelligence suite.