InvDesFlow: AI-Driven Materials Inverse Design

AI-driven materials inverse design service for crystal generation, energy screening, and stability validation

InvDesFlow is a web-based service for AI-driven crystalline materials inverse design, integrating crystal generation, energy screening, and stability validation. Developed by Xiao-Qi Han, Ze-Feng Gao, Peng-Jie Guo, and Zhong-Yi Lu from the School of Physics, Renmin University of China (Lu Group Homepage).

InvDesFlow is currently offered as a free service for academic research. If this tool contributes to your work, we would greatly appreciate citations of our related publications. The information collected below is used only for aggregated usage statistics and will not be made public.

User Information

Example: Renmin University of China - Xiao-Qi Han - hanxiaoqi@ruc.edu.cn

Reference Hull Energy Calculation

CIF file or ZIP archive uMLIP reference-hull model

The service computes candidate uMLIP energies, reuses or fills local MP reference energies, and returns hull-energy CSV files plus a binary or ternary phase-diagram PNG when available.

Download Results

No reference-hull task submitted yet.

After the task is done, download the hull-energy CSV and phase-diagram package here.

Phonon Spectrum Calculation Run full phonon spectrum calculation and dynamical-stability validation through the PhononBench service. Open Service Phonon-aware Scoring Function Rapidly rank candidate materials by phonon-aware dynamical-stability scores before expensive validation. Open Service

User Information

Example: Renmin University of China - Xiao-Qi Han - hanxiaoqi@ruc.edu.cn

Crystal Structure Energy Prediction

CIF file or ZIP archive uMLIP energy model

uMLIP-based PBE energy models predict total_energy_eV and energy_per_atom_eV. A ZIP may contain nested CIF files. More uMLIP models can be added to this selector later.

Download Results

No energy prediction task submitted yet.

After the task is done, download the energy CSV package here.

User Information

Example: Renmin University of China - Xiao-Qi Han - hanxiaoqi@ruc.edu.cn

Crystal Structure Prediction

Chemical formulas Or upload formula TXT Generation model

Use one line per formula as Formula,count; bare formulas default to 10 structures. The service returns a ZIP archive grouped by formula.

Download Results

No structure prediction task submitted yet.

After the task is done, download the generated CIF archive here.

Notes

You may generate candidate structures from formulas, or directly upload your own CIF structures for uMLIP-based PBE energy prediction.
Structure generation options include InvDesFlow-AL for application-oriented inverse-design workflows, DiffCSP is a general formula-conditioned baseline, and DAO provides an additional diffusion-based generator. Generated structures are candidate proposals and should be screened before use.
For the current public crystal-generation service, compact systems are recommended; structures with fewer than about 20 atoms are generally more reliable. Users may use large language models to propose chemical formulas and upload them here, while already-designed candidate structures can be uploaded directly to the screening and validation modules below.
uMLIP options provide fast approximate PBE-level energies for screening and reference-hull comparison. PhononScore is intended for rapid stability ranking, while PhononBench provides full phonon-spectrum validation. DFT should still be used for final high-accuracy decisions.
At most 1000 structures are generated per structure-prediction task. Each formula should request at least 10 structures; larger requests are truncated from top to bottom when the total exceeds the public-service limit.
At most 1000 CIF files are computed per uMLIP energy task. A ZIP archive may contain nested CIF files.
Reference-hull calculation accepts one CIF, or multiple CIF files from the same chemical system, and currently supports unary, binary, and ternary systems with at most 200 CIF files per public-service task.
A mixed-element upload that would require multiple phase diagrams will be rejected; please submit each chemical system as a separate task.
User information is used only for aggregated usage statistics and will not be made public.
For larger-scale screening or collaboration, please contact hanxiaoqi@ruc.edu.cn.

How to Cite

InvDesFlow is an AI-driven materials inverse design service that starts from user-defined chemical formulas or uploaded structures and supports rapid generation, screening, and validation of candidate crystalline materials. The web platform currently connects five connected modules: formula-conditioned crystal structure prediction, uMLIP-based PBE energy prediction, uMLIP-consistent reference-hull calculation, phonon-aware stability scoring, and full phonon-spectrum validation. The overall InvDesFlow methodology is introduced in our review AI-Driven Inverse Design of Materials, and further developed in InvDesFlow and InvDesFlow-AL. If you use InvDesFlow, PhononScore, PhononBench, or related components in your research, please consider citing these works.

@misc{han2026phononscore,
      title={PhononScore: a phonon-aware scoring function for dynamical stability}, 
      author={Xiao-Qi Han and Ze-Feng Gao and Zhong-Yi Lu},
      year={2026},
      eprint={2607.08518},
      archivePrefix={arXiv},
      primaryClass={cond-mat.mtrl-sci},
      url={https://arxiv.org/abs/2607.08518}}
      @misc{han2025phononbench,
    title={PhononBench: A Large-Scale Phonon-Based Benchmark for Dynamical Stability in Crystal Generation},
    author={Xiao-Qi Han and Peng-Jie Guo and Ze-Feng Gao and Zhong-Yi Lu},
    year={2025},
    eprint={2512.21227},
    archivePrefix={arXiv},
    primaryClass={cond-mat.mtrl-sci},
    url={https://arxiv.org/abs/2512.21227}}
@article{InvDesFlow-AL,
    author = {Xiao-Qi Han and Peng-Jie Guo and Ze-Feng Gao and Hao Sun and Zhong-Yi Lu},
    title = {InvDesFlow-AL: active learning-based workflow for inverse design of functional materials},
    journal = {npj Computational Materials},
    year = {2025},
    volume = {11},
    number = {1},
    pages = {364},
    doi = {10.1038/s41524-025-01830-z},
    url = {https://doi.org/10.1038/s41524-025-01830-z},
    issn = {2057-3960},
    date = {2025/11/24}}
@article{InvDesFlow,
    title = {InvDesFlow: An AI-Driven Materials Inverse Design Workflow to Explore Possible High-Temperature Superconductors},
    journal = {Chin. Phys. Lett.},
    volume = {42},
    number = {4},
    pages = {047301},
    year = {2025},
    doi = {10.1088/0256-307X/42/4/047301},
    url = {http://cpl.iphy.ac.cn/en/article/doi/10.1088/0256-307X/42/4/047301},
    author = {Xiao-Qi Han and Zhenfeng Ouyang and Peng-Jie Guo and Hao Sun and Ze-Feng Gao and Zhong-Yi Lu}}
@article{AI4Mreview,
    title = {AI-Driven Inverse Design of Materials: Past, Present, and Future},
    journal = {Chin. Phys. Lett.},
    volume = {42},
    number = {2},
    pages = {027403},
    year = {2025},
    doi = {10.1088/0256-307X/42/2/027403},
    url = {http://cpl.iphy.ac.cn/en/article/doi/10.1088/0256-307X/42/2/027403},
    author = {Xiao-Qi Han and Xin-De Wang and Meng-Yuan Xu and Zhen Feng and Bo-Wen Yao and Peng-Jie Guo and Ze-Feng Gao and Zhong-Yi Lu}}