PaperlayerView exact revision on arXiv

Source-linked AI summary

Property Prediction of Stacked Bilayer Materials: A Multimodal Learning Approach

An Vuong, Minh-Hao Van, Chen Zhao, Xintao Wu

arXiv:2606.01012v1cs.AIcond-mat.mtrl-sci

TL;DR

The paper addresses the limited use of AI for predicting properties of stacked bilayer materials, where diverse configurations and costly simulations create a challenging prediction problem. It proposes BiMat-ML, which jointly encodes monolayer structures, stacking configurations, and monolayer properties without requiring stacked-bilayer CIFs. Experiments report competitive or improved performance across homogeneous and heterogeneous bilayer settings, together with lower training times on HetDB.

  • Problem

    Predicting stacked-bilayer properties is challenging because stacking configurations create a large material space and direct GNNs cannot distinguish intralayer from interlayer interactions.

  • Method

    BiMat-ML jointly models monolayer structures, stacking configurations, and monolayer properties to predict bilayer properties without requiring bilayer CIFs.

  • Results

    BiMat-ML achieves the best overall BiDB bandgap results, with MAE 0.13, MSE 0.07, RMSE 0.26, and R2 = 0.94, while SE-PAINN performs best on HetDB.

  • Takeaways & Limitations

    The framework supports rapid screening and inverse design of stacked 2D materials using separate monolayer information and stacking configurations.

Abstract

from arXiv · show

AI for materials science is a critical topic within AI for science, aiming to accelerate materials discovery and produce accurate property predictions. Bilayer 2D material stacking is essential for exploring new materials with novel functions and inherent phenomena, enabling the creation of new 2D bilayers for diverse real-world applications. Research on bilayer vdWs materials has made significant progress from experimental and computational perspectives. Various bilayer materials have been successfully synthe sized experimentally and the increasing utilization of high-throughput computing technology has con structed several computational two-dimensional materials databases. However, the use of AI to model bilayer stacking and predict new properties remains underexplored, necessitating further research studies. In this work, we propose a novel multimodal learning approach to study the interfaces between dissimilar materials that jointly enable new or multiple functions, and to predict new properties arising from the vertical integration (stacking) of different functional material layers under given configurations. Comprehensive experiments demonstrate the effectiveness and efficiency of our approach compared to baseline methods. Our code is available at https://github.com/AnVuong123/bimat ml.

1 Introduction

Stacking 2D layers enables tunable and emergent properties, but predicting those properties is difficult because stacking configurations create a large design space and direct GNN application conflates intra- and interlayer interactions. BiMat-ML addresses this gap by jointly modeling monolayer structures, stacking configurations, and monolayer properties.

  • Motivation: Bilayer stacking can produce electronic, optical, and mechanical properties absent from individual monolayers, especially under relative twisting that creates moiré superlattices.Stacking is mediated by weak van der Waals forces and can be manipulated to tune material behavior.
  • Challenges: The many possible stacking modes, patterns, and sequences create a large material space whose properties are expensive to evaluate with density functional theory.The computational cost of DFT makes systematic study of stacked 2D materials difficult.
  • Challenges: Directly applying graph neural networks to stacked materials cannot distinguish intralayer from interlayer interactions, limiting accurate property prediction.
  • Proposed approach: BiMat-ML jointly models monolayer structures, stacking configurations, and intrinsic monolayer properties to capture both intra-layer chemistry and interlayer stacking effects.The framework is designed for diverse homobilayer and heterobilayer configurations.

2 Related Work

Prior work establishes that bilayer properties can be tuned through stacking patterns, while AI methods for bilayer 2D materials remain relatively limited. Existing work includes structure-embedded models for stacked-bilayer property prediction, motivating broader multimodal approaches.

  • Bilayer materials: Bilayer 2D materials achieve tunable electronic, optical, and mechanical properties through stacking patterns created by rotations or horizontal layer sliding.
  • AI for materials: AI research has extensively supported materials discovery and prediction tasks, but bilayer 2D materials have received comparatively little attention.The related work identifies structure-embedded PAINN as an existing approach for stacked-bilayer property prediction.

3 Property Prediction of Stacked Bilayer Materials via Multimodal Learning

BiMat-ML predicts bilayer properties from separate monolayer structures, stacking configurations, and monolayer properties, rather than requiring stacked-bilayer CIFs. It encodes crystal graphs and stacking configurations separately, fuses their representations with monolayer properties, and predicts the target property through an MLP.

  • Framework: BiMat-ML takes bottom and top monolayer CIFs, stacking configurations, and known monolayer properties as inputs and outputs a predicted bilayer property.
  • Framework: The framework uses graph encoders for the two monolayers, a stacking configuration encoder, and multimodal fusion to form a joint representation for property prediction.The three components are the graph encoder EG, stacking encoder ES, and fusion module.
  • Monolayer representation: CGCNN constructs crystal graphs from CIF atomic species, coordinates, and periodic neighbors, updates node features with gated convolutions, and mean-pools them into a fixed-dimensional graph representation.Interatomic distances are encoded as edge features using a Gaussian basis expansion.
  • Stacking representation: Stacking configurations affect bilayer behavior through layer sequence, interlayer distance, and twist angle, which alter atomic coordinates and in-plane lattice vectors.
  • Stacking representation: The stacking encoder flattens each stacking matrix, maps it through feed-forward layers into a latent vector, and uses that vector in subsequent multimodal fusion.The encoder is trained to reconstruct the stacking configuration with an L2 loss.

4 Stacking Settings of Bilayer Materials

The paper represents bilayer stacking differently for homobilayers and heterobilayers, encoding transformations and multimodal structural information for property prediction. Heterobilayer construction additionally addresses lattice mismatch through constrained supercell generation and structural optimization.

  • 4.1 Homobilayer Materials: Homobilayer stacking fixes one monolayer and transforms the other using in-plane translation, rotation, and an optional flip.The configuration is represented by an affine transformation matrix whose parameters encode these operations and the interlayer fractional shift.
  • 4.1 Homobilayer Materials: The stacking transformation maps atomic coordinates into the top layer and wraps periodic in-plane coordinates back into the reference unit cell.Under the row-vector convention, the mapping is expressed as Xtop = M(XbottomA), where M applies element-wise modulo-by-1 wrapping.
  • 4.1 Homobilayer Materials: For homobilayers, the model combines a graph representation of the monolayer, an independently encoded stacking configuration, and monolayer properties.These representations are concatenated into a homogeneous bilayer embedding.
  • 4.2 Heterobilayer Materials: Heterobilayer stacking involves relative translation, twist, interlayer distance, and relative sublattice or atomic-species orientation between distinct monolayers.These structures generally have greater stacking complexity than homobilayers.
  • 4.2 Heterobilayer Materials: Because distinct monolayers often have mismatched lattices, the workflow constructs and filters supercell pairs by strain and atom-count constraints before selecting and relaxing an optimal bilayer.Further optimization includes interlayer-distance optimization and structural relaxation to limit computational cost.

5 Experiments

Experiments evaluate BiMat-ML for homobilayer and heterobilayer bandgap prediction, comparing multimodal inputs, baseline models, runtime, and ablations. BiMat-ML is strongest on BiDB and remains competitive without bilayer CIFs on HetDB, while reducing training cost relative to several baselines.

  • Homobilayer Materials: BiMat-ML achieves the best BiDB bandgap prediction, with MAE 0.13, MSE 0.07, RMSE 0.26, and R2 = 0.94.The cited table compares BiMat-ML with baseline models; conventional single-CIF baselines have MAE values around 0.38.
  • Heterobilayer Materials: On HetDB, SE-PAINN performs best with MAE 0.11, RMSE 0.17, and R2 = 0.91, while BiMat-ML reaches MAE 0.13, RMSE 0.21, and R2 = 0.88 without bilayer CIFs.BiMat-ML outperforms Direct on MAE and R2 but trails the strongest structure-based baselines.
  • Heterobilayer Materials: HetDB ablations show that removing monolayer properties causes a larger decline than removing stacking descriptors, while removing both is worst.The weaker correlations between monolayer and bilayer bandgaps on HetDB help explain why these inputs are less critical there than on BiDB.
  • Running Time: BiMat-ML trains in 100 s on HetDB, faster than Direct at 212 s, SE-CGCNN at 236 s, SE-MEGNET at 375 s, and SE-PAINN at 110 s.On BiDB, BiMat-ML and Direct both require 900 s, while SE-CGCNN and SE-MEGNET require 1501 s and 2725 s, respectively.
  • Running Time: All models have 0.5 s per-sample inference time, whereas VASP-based DFT calculations take approximately 4.9 hours.BiMat-ML therefore provides competitive or superior efficiency while avoiding costly DFT-based structure calculation.

6 Conclusion

The paper presents BiMat-ML, a multimodal framework for predicting properties of stacked two-dimensional materials without requiring stacked-bilayer CIFs. Experiments on homogeneous and heterogeneous bilayer datasets demonstrate its effectiveness and efficiency, while the framework remains model-agnostic and extensible.

  • 6 Conclusion: BiMat-ML fuses two monolayer CGCNN encodings with an autoencoder-based stacking representation to predict stacked-bilayer properties without stacked-bilayer CIFs.The framework is evaluated on BiDB for homogeneous bilayers and HetDB for heterogeneous bilayers.
  • 6 Conclusion: Experiments on BiDB and HetDB demonstrate the effectiveness and efficiency of BiMat-ML for stacked-bilayer property prediction.The authors describe the framework as model-agnostic and applicable to a range of graph neural network architectures.
  • 6 Conclusion: The authors plan to investigate other GNN architectures and develop algorithms for identifying stacking configurations that may produce emergent properties.These directions extend the framework beyond its current property-prediction evaluation.

A Example of Stacking Configuration Construction

The Al4S4 example constructs a bilayer stacking configuration from a descriptor and CIF-derived atomic information, then applies the resulting transformation to obtain top-layer fractional coordinates.

  • A Example of Stacking Configuration Construction: The BiDB descriptor specifies a flip transformation, in-plane translation, and stacking configuration for the Al4S4 bilayer.The associated CIF assigns the first four atoms to the bottom monolayer and the remaining four to the top monolayer.
  • A Example of Stacking Configuration Construction: δz = 0.91 is obtained from the Al1/Al3 fractional z coordinates and is uniform across corresponding atom pairs.Using zbottom = 0.28, ztop = 0.63, and p5 = −1 gives δz = 0.63 + 0.28 = 0.91.
  • A Example of Stacking Configuration Construction: The stacking transformation uses p1 = −1, p2 = 0, p3 = 0, p4 = −1, p5 = −1, translation (0.33, −0.33), and δz = 0.91.These parameters are extracted from the CIF-associated construction and combined with atomic information reported in Table 3.
  • A Example of Stacking Configuration Construction: Matrix multiplication maps bottom-layer fractional coordinates Xb through transformation A to produce top-layer coordinates Xt.A modulo-1 operation is applied to in-plane coordinates to enforce periodic boundary conditions.
Loading 2606.01012v1…