Apppublic

sathishbio2/Thermoelectric-Discovery-Platform

sourceHugging Facemitupdated 2mo agoView on Hugging Face

0likes

App README

Thermoelectric Discovery Platform v2.0

AI-guided screening, live prediction, and physics validation for thermoelectric materials discovery

Companion interface for "Fluoride and Oxyfluoride Thermoelectrics Identified by Explainable Machine Learning with 20-Seed Stability Consensus and First-Principles Validation" by Sathish Panneer Selvam et al..

Features

Smart Discovery

AI-guided candidate screening across 10,946 Materials Project compounds. Three modes: Exploratory (high novelty), Validation (proven compositions), and Custom Constraints (user-defined thresholds for stability score, band gap, E_hull, and density).

Live Prediction

Real-time ZT proxy prediction for any user-supplied chemical formula, including compositions not in the training dataset. Uses the trained Random Forest pipeline (134 Magpie features) with prediction uncertainty from the tree ensemble.

SHAP Explanation

Per-compound SHAP feature attribution. Enter any formula and see which Magpie descriptors drive the ZT proxy prediction above or below the dataset mean, with physical interpretation of each contributing feature.

Slack kL Calculator

Lattice thermal conductivity estimation from elastic constants (VASP IBRION=6 output). Includes a dedicated Cahill minimum thermal conductivity mode for mechanically unstable compounds, plus multi-material comparison.

ZT Estimator

Full temperature-dependent ZT(T) from user-supplied Seebeck coefficient, electrical conductivity, and Slack kL. Supports both p-type and n-type carriers with power factor and thermal conductivity breakdown plots.

Similarity Search

Find the 10 most compositionally similar materials to any formula or uploaded CIF file using 132-dimensional Magpie feature cosine similarity.

Dataset Insights

Element frequency analysis, property distributions, and ZT proxy correlations across the full screened database.

BoltzTraP Toolkit

VASP input file generator for band structure, elastic constants, geometry optimization, phonon (DFPT), and DOS calculations. Downloads ready-to-run INCAR/KPOINTS packages.

Export

Download the full dataset, high-stability candidate list, or filtered results as Excel/CSV files.

How It Works

The platform combines pre-computed results from the manuscript's 20-seed Random Forest stability analysis with live, on-demand machine learning inference. Smart Discovery, Similarity Search, and Dataset Insights operate on the pre-screened dataset. Live Prediction and SHAP Explanation run the trained pipeline in real time for any composition the user provides, including formulas outside the original 10,946-compound screen.

The platform enables researchers to:

—Identify robust candidates (materials appearing in top-ranked predictions across at least 95% of 20 random seeds)
—Predict ZT proxy for novel, user-defined compositions
—Understand model predictions through compound-specific SHAP attribution
—Estimate lattice thermal conductivity from elastic constants, including for mechanically unstable phases
—Generate publication-ready ZT(T) curves from BoltzTraP2 transport data
—Produce ready-to-run VASP input files for DFT validation
—Export results for further computational or experimental follow-up

Dataset

—10,946 materials from Materials Project
—Filtered by E_hull <= 0.03 eV/atom, band gap 0.1-2.0 eV
—132 Magpie compositional features plus energyabovehull and n_elements (134 total)
—20-seed stability analysis with Random Forest (R-squared = 0.65)
—10 robust final candidates with stability score >= 0.95

Citation

If you use this tool in your research, please cite:

bibtex

@article{selvam2026thermoelectric,
  title={Fluoride and Oxyfluoride Thermoelectrics Identified by Explainable Machine Learning with 20-Seed Stability Consensus and First-Principles Validation},
  author={Selvam, Sathish Panneer and coauthors},
  journal={},
  year={2026},
  note={Submitted}
}