Cross-Property Transfer Learning and Generative Crystal Design for Accelerated Discovery of Low-Thermal-Conductivity Materials
Harpriya Minhas, Rahul Kumar Sharma, Biswarup Pathak
ACS Applied Materials & Interfaces 2026, 18 (37), 50758–50771
DOI: 10.1021/acsami.6c10770
This repository contains the data, trained-model predictions, and generative-design workflow for the paper above.
Low lattice thermal conductivity (κL) is a key target for thermoelectrics and thermal barrier materials, but labelled κL data are scarce. We address this with a unified framework that combines:
- Graph neural networks (GNNs) — CGCNN, DeeperGATGNN, and ALIGNN benchmarked on 3,925 materials. ALIGNN performs best, as its line-graph (bond-angle) representation captures local bonding and many-body interactions that govern phonon transport.
- Cross-property transfer learning (TL) — models pretrained on six auxiliary properties are fine-tuned for κL. Energy-based source tasks (e.g. formation energy, energy above hull) give the most transferable latent representations for κL prediction.
- Transformer-based generative crystal modelling — CrystaLLM generates 10,000 unique candidate crystal structures, which are screened with the TL models for κL and thermodynamic stability, followed by phonon calculations for dynamical stability.
The pipeline identifies 17 dynamically stable compounds with κL ≤ 0.50 W m⁻¹ K⁻¹.
Curated κL dataset ──► GNN benchmark (CGCNN / DeeperGATGNN / ALIGNN)
│
Auxiliary properties ──► pretrain ─┴─► fine-tune on κL (cross-property TL)
│
CrystaLLM ──► 10,000 generated structures ──► predict κL, E_f, E_hull
│
phonon stability screening
│
17 stable candidates, κL ≤ 0.50 W m⁻¹ K⁻¹
TL_generative_design/
├── Input_data/
│ ├── 3929 selected materials.xlsx # curated κL training set
│ ├── id_prop_new.xlsx # id–property table for GNN training
│ └── crystal_llm_input.xlsx # inputs/prompts for CrystaLLM generation
├── Crystallm/
│ ├── cifs_v1_{train,val,test}.pkl.gz # CIF datasets for CrystaLLM
│ ├── tokens_v1_*.pkl.gz / .tar.gz # tokenised CIF corpora
│ ├── custom_cifs.*, prompts.tar.gz, gen_prompts.tar.gz
│ ├── gen_v1_small_raw.tar.gz # raw generated output
│ ├── generated_cifs.*, generated_processed_cifs.tar.gz
│ ├── clean_raw.py # parse generated CIFs → cleaned CIFs (pymatgen)
│ ├── 3novel_unique_05_26.py # cumulative unique-structure statistics plot
│ ├── unique_structure_statistics.png
│ ├── violin_deviation_plots_final.png
│ └── Generated_poscars/
│ ├── POSCAR_relax1_* # relaxed generated structures
│ └── 2IFC.py # phonon workflow + stability summary (atomate2/jobflow)
├── Results/
│ ├── kappa_pred.xlsx # predicted κL for generated structures
│ ├── FE_pred.xlsx # predicted formation energies
│ └── ehull_pred.xlsx # predicted energy above hull
└── requirements.txt
git clone https://github.com/Harpriyaminhas/TL_generative_design.git
cd TL_generative_design
uv venv && source .venv/bin/activate
uv pip install -r requirements.txtThe GNN models and CrystaLLM are external codes; install them from their own repositories:
- ALIGNN — https://github.com/usnistgov/alignn
- CGCNN — https://github.com/txie-93/cgcnn
- DeeperGATGNN — https://github.com/usccolumbia/deeperGATGNN
- CrystaLLM — https://github.com/lantunes/CrystaLLM
The phonon script (2IFC.py) additionally requires atomate2 and jobflow.
1. Clean generated structures
cd Crystallm
python clean_raw.py # reads ./generated_processed_cif/sample_*.cif → ./cleaned_cifs/2. Unique-structure statistics
python 3novel_unique_05_26.py # edit the input Excel path in the script first3. Phonon stability screening
cd Generated_poscars
python 2IFC.py # runs phonons for POSCAR_relax1_*, writes Results/ and stability_summary.xlsxPaths in the scripts are hardcoded; adjust them to your local layout.
If you use this code or data, please cite:
@article{Minhas2026,
author = {Minhas, Harpriya and Sharma, Rahul Kumar and Pathak, Biswarup},
title = {Cross-Property Transfer Learning and Generative Crystal Design for
Accelerated Discovery of Low-Thermal-Conductivity Materials},
journal = {ACS Appl. Mater. Interfaces},
year = {2026},
volume = {18},
number = {37},
pages = {50758--50771},
doi = {10.1021/acsami.6c10770}
}Please also cite the underlying models (ALIGNN, CGCNN, DeeperGATGNN, CrystaLLM) as appropriate.
Harpriya Minhas — GitHub · Google Scholar