Machine-learning-identified two-dimensional van der Waals multiferroics for four-state nonvolatile memory

By combining machine learning screening with first-principles calculations, this study identifies the AuCrP$_2$S$_6$ monolayer as a promising 2D van der Waals multiferroic candidate that enables non-destructive four-state nonvolatile memory through the intrinsic coupling of its ferroelectric polarization and ferromagnetic order via the bulk photovoltaic effect.

The Gist

Imagine you are trying to build a super-advanced, ultra-dense library where every single book is the size of a single atom. In this library, you don't just want to store "Yes" or "No" (0 or 1); you want to store four different states at once (00, 01, 10, 11) to pack twice as much information into the same tiny space.

To do this, the scientists in this paper looked for a special kind of "magic material" that has two superpowers at the same time:

1. Electric Switching: It can flip its electrical charge direction (like a magnet, but for electricity).
2. Magnetic Switching: It can flip its magnetic direction.

Usually, finding a material that does both is like finding a unicorn; they are incredibly rare because the rules of physics that make electricity switch often fight against the rules that make magnetism work.

The Search: A Digital Detective Story

Since these materials are so rare, the researchers didn't just guess. They used a Machine Learning "Detective" to sift through thousands of possible chemical combinations.

Think of the chemical world as a massive, messy attic filled with millions of boxes. Most boxes are empty or contain junk (materials that can't be built). A few contain the "treasure" (materials that can be built). The problem is, the detective only has a list of a few known treasures, but no list of the junk.

To solve this, the team taught their AI a special trick called "PU-Bagging." Instead of guessing that every unknown box is junk, the AI plays a game of "what if?" It pretends different groups of unknown boxes are junk, trains itself, and then combines all those guesses to create a confidence score. It's like asking a hundred different detectives to look at the attic and voting on which boxes are most likely to contain treasure.

They also used Transfer Learning, which is like teaching the AI to recognize 3D buildings (bulk crystals) first, and then teaching it how to recognize 2D "flat sheets" (monolayers) based on what it already knows. This helped them find the best candidates even though there wasn't much data on 2D materials to begin with.

The Discovery: The Gold-Crystal-Sulfur Sheet

After the AI narrowed down the list, the researchers used super-computers to simulate the top candidates. They found a winner: a single layer of atoms made of Gold (Au), Chromium (Cr), Phosphorus (P), and Sulfur (S).

Think of this material as a tiny, flexible trampoline made of atoms:

• The Magnetism: The Chromium atoms act like tiny compass needles that all point in the same direction.
• The Electricity: The Gold atoms can slide up and down on this trampoline. When they slide to one side, the material becomes electrically positive on top and negative on the bottom. When they slide to the other side, it flips.
• The Stability: The Gold atoms can flip back and forth easily (like a light switch) without getting stuck, but they stay put once you let go (non-volatile memory).

The Reading Trick: The "Light Flash"

The biggest problem with these memory devices is usually how to read the information without breaking it. Traditional methods often zap the material, erasing the data before you can read it.

The researchers found a clever way to read the data using light, specifically a phenomenon called the Bulk Photovoltaic Effect (BPVE). Imagine shining a flashlight on the material:

1. The Electric Signal: Depending on which way the Gold atoms are shifted (the electric state), the light will push electrons to flow either Left or Right. This tells you the "0" or "1" of the electric bit.
2. The Magnetic Signal: Because the material is magnetic, it acts like a bouncer at a club. It only lets electrons with a specific "spin" (a quantum property, like a tiny top spinning clockwise or counter-clockwise) pass through. If the magnetic field points one way, only "clockwise" electrons flow. If it flips, only "counter-clockwise" electrons flow.

The Result: A Four-State Memory Cell

By combining these two signals, the material can store four distinct states in a single atomic layer:

• State 00: Electric Left + Clockwise Spin
• State 01: Electric Left + Counter-Clockwise Spin
• State 10: Electric Right + Clockwise Spin
• State 11: Electric Right + Counter-Clockwise Spin

The scientists propose a device where you write data by flipping the electric or magnetic switches, and you read it by shining a light and measuring the direction of the current and the type of spin. This allows for a non-destructive readout, meaning you can check the memory without erasing it.

In short, this paper presents a blueprint for a new type of computer memory that is twice as dense as current technology, found using a smart AI detective, and read using a clever light-based trick.

Technical Summary

Technical Summary: Machine-Learning-Identified Two-Dimensional van der Waals Multiferroics for Four-State Nonvolatile Memory

Problem Statement
The development of high-density nonvolatile memory in the post-Moore era requires materials that integrate ferroelectric (FE) and magnetic orders within a single system. While two-dimensional (2D) van der Waals (vdW) multiferroics offer a promising platform for four-state memory (encoding logic via distinct polarization $P$ and magnetization $M$ configurations), their practical realization is hindered by two primary challenges:

1. Scarcity of Candidates: Intrinsically multiferroic 2D monolayers are exceptionally rare due to the fundamental electronic incompatibility between ferroelectricity and magnetism.
2. Readout Limitations: Conventional electrical detection of FE states is often destructive, requiring re-writing cycles. Furthermore, distinguishing coupled $(P, M)$ states is frequently compromised by signal crosstalk.
3. Data Bias in Discovery: Traditional machine learning (ML) approaches struggle with the "positive-unlabeled" (PU) data bias inherent in materials science, where experimentally synthesized materials provide reliable positive examples, but the vast majority of hypothetical structures are merely unlabeled rather than definitive negatives for synthesizability.

Methodology
The authors propose a hybrid framework combining machine-learning screening with first-principles calculations to explore the chemically versatile $ABC_2X_6$ vdW family (where $A/B$ are cations, $C$ is a pnictogen, and $X$ is a chalcogen).

• Machine Learning Screening:

  • PU-Bagging Strategy: To address data scarcity and bias, the authors employ a semi-supervised PU-bagging strategy using Crystal Graph Convolutional Neural Networks (CGCNNs). Instead of treating unlabeled structures as true negatives, the model iteratively resamples pseudo-negative subsets to train multiple sub-classifiers, aggregating their predictions into a final confidence-level (CL) score.
  • Transfer Learning: To bridge the gap between sparse 2D data and abundant 3D data, the model is pre-trained on large-scale 3D bulk crystal entries from the Materials Project (validated against the ICSD) and then fine-tuned on the target 2D $ABC_2X_6$ dataset using parameter regularization.
  • Feature Analysis: A gradient boosting (GB) regressor, identified as the most accurate model for predicting polarization, is used to screen for large out-of-plane polarization. Shapley additive explanations (SHAP) are employed to interpret feature importance, revealing that mean atomic radius and maximum metal period are key determinants.

• First-Principles Validation:

  • High-confidence candidates (CL score $\ge$ 0.9) undergo density functional theory (DFT) calculations to verify ferroelectricity, magnetism, and switching barriers.
  • Nonlinear optical properties, specifically the bulk photovoltaic effect (BPVE) and shift current, are calculated using maximally localized Wannier functions to assess non-destructive readout capabilities.

Key Results
The screening process identified 739 high-confidence candidates from the $ABC_2X_6$ family. Among these, monolayer AuCrP$_2$S$_6$ emerged as a representative multiferroic semiconductor with the following properties:

• Multiferroic Nature: It possesses a ferromagnetic ground state in both its ferroelectric (FE) and paraelectric (PE) phases.
• Ferroelectric Properties: The material exhibits a sizable out-of-plane spontaneous polarization of 7.46 pC/m. The switching barrier is moderate at ~130 meV/f.u., suggesting robustness against thermal fluctuations while remaining switchable under practical electric fields.
• Electronic Structure: The FE phase is a direct bandgap semiconductor (1.06 eV at the K point) with significant exchange splitting. The PE phase has an indirect gap of 0.74 eV, yet retains the magnetic exchange splitting.
• Non-Destructive Readout Mechanism:
  • Polarization Encoding: The shift current (a nonlinear optical response) is highly sensitive to polarization direction. Reversing $P$ deterministically reverses the sign of the photocurrent ($\sigma_{xyy}$) while maintaining magnitude.
  • Magnetization Encoding: Due to robust exchange splitting, the photocurrent is spin-selective. The active spin channel (spin-up or spin-down) is uniquely locked to the magnetic order of the Cr atoms.
  • Dual-Channel Probe: This intrinsic coupling allows for the simultaneous decoding of $P$ (via current direction) and $M$ (via spin channel), enabling the non-destructive readout of four distinct logic states: $(+P, +M)$, $(+P, -M)$, $(-P, +M)$, and $(-P, -M)$.

Significance and Claims
The paper claims to provide a practical blueprint for next-generation multistate optoelectronic memory. Specifically:

1. Methodological Advancement: The study establishes a generalized paradigm for data-sparse screening by combining PU-bagging and transfer learning, offering a scalable approach to discover functional materials where experimental data is limited.
2. Material Discovery: It identifies AuCrP$_2$S$_6$ as a high-confidence candidate that overcomes the rarity of intrinsic 2D multiferroics, offering a large polarization and a moderate switching barrier.
3. Device Concept: The authors propose a unified optoelectronic readout scheme for a four-state (2-bit per cell) nonvolatile memory device. By exploiting the polarization-controlled photocurrent direction and the magnetization-selected active spin channel, the device enables deterministic, non-destructive decoding of all four logic states within a single atomic layer.

The work positions 2D multiferroics as a viable solution for high-density memory in the post-Moore era, moving beyond simple material identification to propose a functional mechanism for state readout that avoids the destructive limitations of conventional electrical detection.