strain2DOS: Machine Learning Reconstruction of Local Electronic Structure in Non-Uniformly Strained MoS₂

strain2DOS is an open-source project developed by SHazra321 (GitHub link: https://github.com/SHazra321/strain2DOS). Its core is to predict the density of states (DOS) distribution of molybdenum disulfide (MoS₂) from Raman spectroscopy data using machine learning, enabling fast reconstruction of the local electronic structure of materials under non-uniform strain. The corresponding paper for this project is arXiv:2603.29298, which addresses the issues of high cost and difficulty in large-scale spatial mapping of traditional first-principles calculations (DFT), providing an efficient tool for two-dimensional material research.

The 2D material MoS₂ is widely used in flexible electronics, optoelectronic devices, and other fields. However, mechanical strain changes its electronic structure, affecting device performance. While traditional DFT calculations are accurate, they are costly and cannot achieve large-scale spatial mapping. The strain2DOS project aims to use data-driven methods to replace expensive DFT calculations and quickly obtain local electronic structure information of MoS₂ under strained conditions.

The project uses supervised learning to establish the mapping relationship between Raman spectral features and DOS. The process is as follows:

1. Data Collection: Obtain DOS data under different strain configurations via DFT calculations;
2. Feature Engineering: Extract strain-related features from Raman spectra;
3. Model Training: Use neural networks to learn the mapping from spectral features to DOS;
4. Spatial Reconstruction: Apply the model to experimental Raman strain maps to reconstruct local electronic structures. Neural networks need to address challenges such as multi-input mapping, physical constraints (e.g., non-negativity), and spatial generalization.

The project provides two core Jupyter Notebooks:

• Training_Prediction_DOS.ipynb: Responsible for model training and basic prediction, including steps like loading DFT training data (train_dos.csv), building the model, training monitoring, and evaluation;
• PostProcessing_DOS_map.ipynb: Used for post-processing and spatial mapping, reading model weights and Raman strain data (raman_strain.csv), and outputting spatially resolved DOS distribution maps. Dataset Description: train_dos.csv contains DOS training data from DFT calculations, and raman_strain.csv is sample strain map data.

System Requirements:

Item	Minimum Requirement	Recommended Configuration
Python	3.8+	3.10+
Memory	8GB	16GB+
CPU	Multi-core processor	Modern multi-core CPU
GPU	Optional	NVIDIA GPU (for training acceleration)

Installation Steps:

1. Clone the repository: git clone https://github.com/SHazra321/strain2DOS.git && cd strain2DOS
2. Install dependencies: pip install -r requirements.txt && pip install jupyter
3. Launch Jupyter: jupyter notebook After running Training_Prediction_DOS.ipynb, run PostProcessing_DOS_map.ipynb to complete spatial reconstruction.

After successful execution, the following visualization results can be obtained:

• Training loss curve: Shows the model convergence process;
• Bandgap prediction plot: Comparison between model predictions and DFT calculation values;
• Spatial bandgap reconstruction map: 2D spatial bandgap distribution heatmap;
• Experimental PL data comparison: Validation with experimental photoluminescence data. These results help understand the impact of strain on the electronic properties of MoS₂ and guide device design.

The scientific significance and application value of the project include:

1. Accelerated Material Characterization: Replace expensive DFT calculations to quickly obtain electronic structure information and speed up the characterization process;
2. Non-Destructive Testing Potential: Combine the non-destructive nature of Raman spectroscopy to enable real-time monitoring of electronic structures in operating devices;
3. Methodological Universality: Although focused on MoS₂, the idea of spectral-electronic structure mapping can be extended to other 2D material systems, providing a new paradigm for materials informatics.

Limitations:

• Data Dependence: Model performance is limited by the coverage of training data;
• Physical Interpretability: The black-box nature of neural networks makes it difficult to understand physical mechanisms;
• Generalization Ability: May perform poorly on strain modes outside the training set.

Future Directions:

• Introduce physically constrained loss functions to improve the physical rationality of prediction results;
• Use explainable AI techniques to reveal the physical mechanisms of spectral-DOS mapping;
• Extend to complex material systems such as multi-layer heterojunctions.