Decoding Non-Linearity and Complexity: Deep Tabular Learning Approaches for Materials Science
DataScribe Deep Tabular Learning This repository provides the code, data, and documentation for the paper: Decoding Non-Linearity and Complexity: Deep Tabular Learning Approaches for Materials Science Vahid Attari, Raymundo Arroyave, Department of Materials Science & Engineering, Texas A&M University. DataScribe GUI demo showing encoder-decoder training on HEA tabular data. Watch on YouTube 📖 Overview This project addresses the challenges in tabular materials science data, including extreme skewness, and complex feature interactions. The repository implements several machine learning models, ranging from classical tree-based methods (e.g., XG
Resource information
- Format
- zip
- Original repository
- 16396374
- DOI
- 10.5281/zenodo.16396374
- Credit basis
- Creators in Zenodo record metadata
- Access
- Open files available from the original Zenodo deposit
- Last checked
- Nov 5, 2026
- License
- cc-by-4.0
- Last record update
- Nov 5, 2026
About this resource
DataScribe Deep Tabular Learning This repository provides the code, data, and documentation for the paper: Decoding Non-Linearity and Complexity: Deep Tabular Learning Approaches for Materials Science Vahid Attari, Raymundo Arroyave, Department of Materials Science & Engineering, Texas A&M University. DataScribe GUI demo showing encoder-decoder training on HEA tabular data. Watch on YouTube 📖 Overview This project addresses the challenges in tabular materials science data, including extreme skewness, and complex feature interactions. The repository implements several machine learning models, ranging from classical tree-based methods (e.g., XGBoost) to advanced deep learning architectures (e.g., encoder-decoder models, TabNet, and Variational Autoencoders) on various materials science datasets. Key features of the project: Tabular data preprocessing for skewness and outliers. Benchmarking of deep learning models (e.g., regularized Dense Networks, Disjunctive Normal Form Networks, TabNet). Bayesian optimization for hyperparameter tuning. Comparative analysis of predictive performance and computational efficiency. 📊 Datasets ATLAS-RHEA Description The dataset includes tabular materials data with features such as: Compositional information : Nb, Cr, V, W, Zr. Material properties : Thermal conductivity, density, yield strength, creep rate. BIRDSHOT Description The dataset includes tabular materials data with features such as: Compositional information : Al Co Cr Cu Fe Mn Ni V (Bayesian and EDS measured). Material properties : Computed and experimentally evaluated mechanical properties including yield and ultimate tensile strength, hardness, depth of penetration, SFE, etc. BORG-HEA Description The dataset includes tabular materials data with features such as: Compositional information : Multiple elements Material properties : Mechanical properties Preprocessing Missing values are handled via interpolation. Features are normalized using quantile transformation for skewed distributions. Models The repository implements and benchmarks the following models: Encoder-Decoder (asymmetric overcomplete) regularized Dense Neural Network (rDN) : Standard regularized dense architecture. Encoder-Decoder Disjunctive Normal Form Network (DNF-Net) : Captures logical relationships in tabular data. Encoder-Decoder (1D-CNN) : Extracts local patterns and hierarchical feature representations using 1D convolutional filters, suitable for structured tabular data with implicit ordering or grouped feature patterns.. TabNet : Attention-based model that performs sequential feature selection, enabling interpretability while capturing complex feature interactions in tabular data. Variational Autoencoder (VAE) : Generative model for handling uncertainty. XGBoost : Baseline tree-based model. Key Metrics Mean Squared Error (MSE) Mean Squared Logarithmic Error (MSLE) R-squared (R²) Symmetric Mean Absolute Percentage Error (SMAPE)