🔬 mesoscale physics

Machine-Learned Hamiltonians for Quantum Transport Simulation of Valence Change Memories

This paper introduces an equivariant graph neural network approach that accurately predicts Hamiltonian matrices for large, non-periodic valence change memory systems containing thousands of atoms, thereby overcoming the computational and memory limitations of traditional density-functional theory to enable quantum transport simulations of large-scale devices.

Original authors: Chen Hao Xia, Manasa Kaniselvan, Marko Mladenoivić, Mathieu Luisier

Published 2026-02-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Chen Hao Xia, Manasa Kaniselvan, Marko Mladenoivić, Mathieu Luisier

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand how electricity flows through a tiny, complex machine made of atoms. To do this accurately, scientists use a powerful mathematical tool called Density-Functional Theory (DFT). Think of DFT as a super-precise, high-resolution camera that takes a picture of every single atom and calculates exactly how they interact.

However, there's a catch: this "camera" is incredibly slow and expensive to run. If your machine is small and neat (like a perfect crystal), the camera works fine. But if your machine is messy, disordered, or broken—like the Valence Change Memory (VCM) devices used in next-generation computers—the camera has to zoom out to see thousands of atoms at once. At that size, the calculation takes so long and requires so much computer memory that it becomes impossible to finish.

The Solution: A "Smart Apprentice"

The authors of this paper built a Machine Learning (ML) apprentice to solve this problem. Instead of asking the slow, expensive "camera" (DFT) to do all the work, they taught a fast, smart AI to guess the answers.

Here is how they did it, using some everyday analogies:

1. Learning the Rules of the Game
The "answer" the scientists need is a giant grid of numbers called a Hamiltonian matrix. This grid is like a map of the electrical connections between every atom in the device.

The Problem: In messy materials (like the amorphous oxides inside memory chips), the atoms aren't lined up in neat rows. They are scattered randomly, making the map very hard to draw.
The AI's Job: The team trained an Equivariant Graph Neural Network (EGNN). Imagine this network as a detective who learns the "rules of the road" for atoms. It learns that if you rotate the whole device, the electrical map should rotate with it, not change completely. This allows the AI to learn from very few examples and apply those rules to huge, messy structures it has never seen before.

2. The "Augmented Partitioning" Trick
The memory devices they studied are huge (containing over 5,000 atoms), but the AI's "brain" (the computer's memory) is too small to hold the whole map at once.

The Analogy: Imagine trying to read a massive encyclopedia, but you can only hold one page at a time.
The Solution: The researchers used a technique called augmented partitioning. They sliced the giant device into thin, manageable layers (like slicing a loaf of bread). The AI reads one slice, but it also keeps a "note" about the atoms in the neighboring slices so the connections aren't broken. This lets the AI reconstruct the full map piece by piece without running out of memory.

3. The Results: Fast and Mostly Accurate
The team tested their AI on a memory device made of Titanium Nitride and Hafnium Oxide.

Speed: The AI predicted the electrical map in 2 seconds. The traditional method (DFT) would have taken nearly 4 hours on a supercomputer.
Accuracy: The AI's map was very close to the "perfect" map. The errors were tiny (about the size of a grain of sand compared to a mountain).
The Catch: While the map was very accurate, the final result—the electric current flowing through the device—was only "qualitatively" good.
- Analogy: Imagine the AI drew a map of a city's roads with 99% accuracy. If you drive a car using that map, you will generally get to the right neighborhood, but you might miss a specific turn or hit a small bump. The AI correctly predicted whether the device was "on" (conducting electricity well) or "off" (blocking electricity), but the exact number of electrons flowing wasn't perfect.

Why This Matters

The paper claims that this approach allows scientists to study huge, messy devices that were previously too big to simulate. By swapping the slow "camera" for the fast "apprentice," they can now simulate how these memory devices change over time (like when a conductive path forms or breaks) without waiting days for results.

The authors suggest this could be a stepping stone to studying even more complex devices, like phase-change memories (which switch between solid and liquid states), but they stop short of claiming it is ready for commercial use or medical applications. They emphasize that while the speed is a huge win, the accuracy still needs a little more polishing to be perfect.

Technical Summary: Machine-Learned Hamiltonians for Quantum Transport Simulation of Valence Change Memories

Problem Statement
Ab-initio device simulations based on Density-Functional Theory (DFT) are essential for capturing atomistic details in modern electronic devices. However, constructing the Hamiltonian matrix ( $H$ ) is computationally expensive, scaling as $O(N^3)$ where $N$ is the number of atoms. While periodic boundary conditions allow for efficient calculations in homogeneous structures, this approach fails for amorphous or defective materials, such as the switching layers in Resistive Random Access Memories (ReRAM). Accurately modeling these disordered systems requires large domains containing thousands of atoms, pushing standard DFT tools beyond their computational and memory limits. Consequently, investigating the "current vs. voltage" characteristics of evolving device morphologies, such as the formation and dissolution of conductive filaments in Valence Change Memories (VCM), becomes prohibitively difficult.

Methodology
To overcome these bottlenecks, the authors propose replacing DFT-computed Hamiltonians with machine-learned (ML) predictions using an Equivariant Graph Neural Network (EGNN).

Network Architecture: The EGNN is designed to handle the rotational covariance inherent in electronic structure predictions. It utilizes a strictly local, single message-passing layer where nodes and edges represent onsite and offsite Hamiltonian blocks, respectively. The network employs multi-headed attention to distinguish complex atomic environments and enforces physical constraints regarding geometric rotations.
Training Strategy: The model is trained on the Hamiltonian matrices of two TiN- $HfO_2$ -Ti/TiN VCM structures containing uniformly distributed oxygen vacancies. Crucially, these training samples were generated by randomly inserting vacancies into stoichiometric layers rather than via Kinetic Monte Carlo (KMC), creating a significant distributional shift between training and test data. This setup rigorously tests the model's ability to generalize to unseen, physically meaningful vacancy configurations (clustered vacancies and filament morphologies).
Augmented Partitioning: To handle large structures (~5,000 atoms) within the memory constraints of a single GPU, the authors employ an "augmented partitioning" technique. The device is divided into longitudinal slices (in the $x-y$ plane) to capture material interfaces (TiN, Ti, $HfO_2$ ), while ensuring atomic connectivity across partition boundaries is fully accounted for to maintain prediction accuracy.
Simulation Workflow: Once trained, the EGNN predicts the Hamiltonian matrix ( $H_{pred}$ ) for full device structures with various filament states (formed or broken). These predicted matrices are fed into an in-house quantum transport simulator using the Non-Equilibrium Green's Function (NEGF) formalism to compute the energy-resolved transmission function $T(E)$ and electrical current ( $I_d$ ) at a 1 V bias.

Key Results
The study was validated on VCM cells containing 5,268 atoms with four distinct configurations: two with fully formed filaments and two with broken filaments.

Hamiltonian Prediction Accuracy: The model achieved a Mean Absolute Error (MAE) of 3.39 to 3.58 meV when predicting Hamiltonian matrix entries compared to DFT references. The node and edge errors were consistently low (1.54–1.82 m $E_h$ for nodes and ~0.12 m $E_h$ for edges) across all test cases, with minimal outliers. This performance is comparable to state-of-the-art results (2.2 meV) but applied to structures orders of magnitude larger (5,268 atoms vs. $\le$ 150 atoms).
Transport Properties: While the predicted Hamiltonian entries were highly accurate, the resulting energy-resolved transmission functions $T(E)$ showed only qualitative agreement with DFT. Although band edges were well-reproduced, specific features like transmission peaks were not accurately captured.
Current Prediction: Despite the limitations in $T(E)$ , the calculated electrical currents were sufficiently close to DFT counterparts to correctly assess the conductance state of the device. The model demonstrated higher accuracy for configurations with fully formed filaments compared to broken filaments, where low transmission values are more sensitive to Hamiltonian errors.
Computational Efficiency: The ML approach offers a dramatic speedup, requiring only 2 seconds for a forward pass compared to 3.94 node-hours for DFT. This efficiency allows for the rapid construction of Hamiltonians for hundreds of intermediate samples along a resistance transition, amortizing the initial training cost (<40 node-hours).

Significance and Claims
The authors claim that this work provides a viable path to overcome the memory and computational limits of DFT, enabling the study of large-scale, disordered devices that are currently beyond ab-initio capabilities. The primary contribution is demonstrating that an EGNN, trained on a limited set of simple configurations, can generalize to predict Hamiltonians for complex, large-scale VCM structures with unseen vacancy distributions.

The paper modestly concludes that while the current ML model facilitates the investigation of evolving device morphologies by bypassing intensive DFT calculations, further enhancements in model accuracy are required to fully replace DFT for all transport properties. The authors envision future applications in other memory technologies, such as phase-change memories, where gradual amorphous-to-crystalline transitions could similarly be modeled using ML instead of DFT.

The Solution: A "Smart Apprentice"

Why This Matters

More like this