Physics-Guided Sequence Modeling for Fast Simulation and Design Exploration of 2D Memristive Devices

TL;DR

This paper presents a physics-guided neural network framework that accelerates the simulation and design of 2D memristive devices by combining high-fidelity models with machine learning, enabling rapid, accurate, and interpretable predictions.

Contribution

The authors develop a hybrid modeling approach integrating finite-volume simulations with an LSTM neural network for fast, accurate, and physically interpretable predictions of 2D memristive device behavior.

Findings

Achieves over 10,000x speedup compared to traditional simulations

Maintains <1% normalized error in predictions

Enables efficient design exploration and inverse modeling

Abstract

Modeling hysteretic switching dynamics in memristive devices is computationally demanding due to coupled ionic and electronic transport processes. This challenge is particularly relevant for emerging two-dimensional (2D) devices, which feature high-dimensional design spaces that remain largely unexplored. We introduce a physics-guided modeling framework that integrates high-fidelity finite-volume (FV) charge transport simulations with a long short-term memory (LSTM) artificial neural network (ANN) to predict dynamic current-voltage behavior. Trained on physically grounded simulation data, the ANN surrogate achieves more than four orders of magnitude speedup compared to the FV model, while maintaining direct access to physically meaningful input parameters and high accuracy with typical normalized errors <1%. This enables iterative tasks that were previously computationally prohibitive, including inverse modeling from experimental data, design space exploration via metric mapping and sensitivity analysis, as well as constrained multi-objective design optimization. Importantly, the framework preserves physical interpretability via access to detailed spatial dynamics, including carrier densities, vacancy distributions, and electrostatic potentials, through a direct link to the underlying FV model. Our approach establishes a scalable framework for efficient exploration, interpretation, and model-driven design of emerging 2D memristive and neuromorphic devices.