An open platform for interactive benchmarking and simulations using machine learning interatomic potentials
MLIP Studio is a Streamlit-based web application for running, testing, comparing, and benchmarking universal machine learning interatomic potentials (MLIPs) for atomistic simulations of molecules and materials. It brings more than 60 pretrained MLIPs into a single graphical interface, so users can evaluate structures, compare models, run common simulation tasks, and benchmark against reference data without writing custom scripts for every model family.
The platform is designed for computational chemistry, materials science, MLIP model development, dataset diagnostics, and teaching. It combines model inference, 3D visualization, geometry optimization, trajectory analysis, and benchmarking tools in one workflow.
MLIP Studio can be tried online for free at:
The hosted instance provides GPU acceleration and is intended for quick tests, demonstrations, education, and small-to-moderate systems. Because it is a shared public service, it has limits on system size and runtime.
For larger systems, unrestricted benchmarking, long calculations, or private datasets, users can download and install MLIP Studio locally. Local installation removes the cloud system-size limits and lets the application use the user's own CPU/GPU hardware.
Universal MLIPs are foundation models for atomistic simulation. They are trained on large quantum-mechanical datasets and can often be applied across broad chemical spaces without system-specific retraining. In practice, however, using several MLIP families side by side can be difficult because each model stack may require different dependencies, APIs, licenses, and data conventions.
MLIP Studio addresses this by providing a unified interface for model selection, structure input, calculation setup, result visualization, and benchmark analysis. Users can switch between model families in one session and compare predictions on the same molecules, crystals, slabs, interfaces, or trajectory datasets.
| Area | Capabilities |
|---|---|
| Structure input | Built-in examples, file upload, pasted structure text, Materials Project ID import, PubChem import, batch upload, and extXYZ trajectory upload |
| Supported formats | XYZ, extXYZ, CIF, POSCAR/VASP, CONTCAR, MOL, SDF, and Turbomole-style molecular files, through ASE and related parsers |
| Visualization | Interactive 3D structure viewing with multiple styles such as ball-and-stick, stick, line, and space-filling representations |
| Model selection | Predefined universal MLIPs, custom MACE model upload, custom MACE model URL loading, UFF, D3 dispersion, xTB, and the in-house QM9 HOMO-LUMO gap model |
| Single-structure prediction | Energy, forces, stress, atomization energy, cohesive energy, HOMO-LUMO gap, band gap, density of states, dipole moment, partial charges, and Hessian-related outputs where supported |
| Geometry optimization | Molecular geometry optimization and cell plus geometry optimization using ASE optimizers including BFGS, BFGSLineSearch, LBFGS, LBFGSLineSearch, FIRE, GPMin, MDMin, and FASTMSO |
| FASTMSO optimizer | A multi-stage optimizer that switches through FIRE, MDMin, and LBFGS for robust MLIP-based relaxation |
| Vibrational analysis | Finite-difference vibrational mode analysis, frequency tables, frequency histograms, zero-point vibrational energy, vibrational entropy, and downloadable CSV output |
| Equation of state | Energy-volume scans with Birch-Murnaghan, Murnaghan, and Vinet fits; reports equilibrium volume, equilibrium energy, bulk modulus, and pressure derivative |
| Spin-state scans | Spin determination for compatible OMOL-style models that accept charge and spin inputs |
| Batch processing | Multi-file batch evaluation of energy, forces, and stress; batch atomization/cohesive energy; batch HOMO-LUMO gap prediction |
| Trajectory benchmarking | extXYZ trajectory evaluation with parity plots, error tables, MAE, RMSE, R2 metrics, element-wise force diagnostics, and downloadable results |
| Performance benchmarking | Wall-clock time reporting for comparing model speed across CPU/GPU hardware and model families |
| Deployment | Local Streamlit app, hosted online instance, and a provided Dockerfile for containerized deployment |
MLIP Studio currently includes 62 predefined universal MLIP models from six major model families. Additional calculators and task-specific models are also available.
| Family | Count | Models |
|---|---|---|
| MACE | 24 | MACE MPA Medium MACE OMAT Medium MACE OMAT Small MACE MATPES r2SCAN Medium MACE MATPES PBE Medium MACE MP 0a Small MACE MP 0a Medium MACE MP 0a Large MACE MP 0b Small MACE MP 0b Medium MACE MP 0b2 Small MACE MP 0b2 Medium MACE MP 0b2 Large MACE MP 0b3 Medium MACE ANI-CC Large (500k) MACE OMOL-0 XL 4M MACE OMOL-0 XL 1024 MACE OFF 23 Large MACE OFF 23 Medium MACE OFF 23 Small MACE OFF 24 Medium MACE POLAR 1 S MACE POLAR 1 M MACE POLAR 1 L |
| FAIRChem | 5 | UMA Small 1.2 UMA Small 1.1 ESEN MD Direct All OMOL ESEN SM Conserving All OMOL ESEN SM Direct All OMOL |
| ORB v3 | 10 | V3 OMOL Conservative V3 OMOL Direct V3 OMAT Conservative (inf) V3 OMAT Conservative (20) V3 OMAT Direct (inf) V3 OMAT Direct (20) V3 MPA Conservative (inf) V3 MPA Conservative (20) V3 MPA Direct (inf) V3 MPA Direct (20) |
| MatterSim | 2 | V1 SMALL V1 LARGE |
| PET / UPET | 16 | PET-MAD-XS-V1.5.0 PET-MAD-S-V1.5.0 PET-MAD-S-V1.1.0 PET-MAD-S-V1.0.2 PET-OAM-L-V0.1.0 PET-OMAT-XS-V1.0.0 PET-OMAT-S-V1.0.0 PET-OMAT-M-V1.0.0 PET-OMAT-L-V1.0.0 PET-OMATPES-L-V0.1.0 PET-SPICE-S-V0.2.0 PET-SPICE-L-V0.2.0 PET-MAD-DOS PET-OMAD-XS-V1.0.0 PET-OMAD-S-V1.0.0 PET-OMAD-L-V0.1.0 |
| SevenNet | 5 | 7net-0 7net-l3i5 7net-omat 7net-mf-ompa 7net-omni |
| Calculator or model | Use |
|---|---|
| xTB | Semi-empirical tight-binding calculations through the external xtb executable. xTB must be preinstalled and available on the system PATH. |
| UFF | Classical Universal Force Field calculator, currently recommended mainly for energy evaluation. |
| D3 dispersion | Standalone DFT-D2/DFT-D3 style dispersion correction through torch-dftd. |
| In-house QM9 gap model | Lightweight message-passing model for HOMO-LUMO gap prediction of QM9-like organic molecules. |
- Load a molecule, crystal, surface, interface, or trajectory from an example, local file, pasted text, PubChem, Materials Project, batch upload, or extXYZ trajectory.
- Inspect the structure in the interactive 3D viewer.
- Select a model family and a specific pretrained model.
- Choose CPU or CUDA GPU execution when available.
- Run a calculation task such as energy/force/stress evaluation, geometry optimization, vibrational analysis, equation of state fitting, spin-state determination, or batch benchmarking.
- View plots, tables, parity plots, error metrics, optimized structures, and downloadable output files in the browser.
- Python 3.10 is recommended.
- Git is required for installing several model packages from source.
- A working C/C++ build environment may be needed by some scientific Python dependencies.
- xTB must be installed separately and available as
xtbon the systemPATHif you want to use the xTB calculator. - A CUDA-capable GPU and compatible PyTorch installation are required for local GPU acceleration.
- Materials Project import requires an
MP_API_KEYenvironment variable. - UMA and ESEN models also require a hugging face login and approval of the account for downloading UMA and ESEN models.
Clone the repository and install the dependencies:
git clone <repository-url>
cd <repository-name>
python -m venv .venvActivate the environment.
On Linux or macOS:
source .venv/bin/activateOn Windows PowerShell:
.\.venv\Scripts\Activate.ps1Install dependencies and run the app:
pip install --upgrade pip
pip install --no-deps fairchem-core==2.16.0 --ignore-requires-python
pip install -r requirements.txt
pip install --no-deps "upet@git+https://github.com/lab-cosmo/upet.git" --ignore-requires-python
pip install --no-deps "git+https://github.com/WillBaldwin0/graph_electrostatics.git" --ignore-requires-python
streamlit run Home.pyThese commands mirror the dependency installation order used in the provided Dockerfile.
The app will start locally at the URL printed by Streamlit, usually:
http://localhost:8501
A Dockerfile is provided for users who prefer a containerized setup or want a reproducible deployment recipe.
docker build -t mlip-studio .
docker run --rm -p 8501:8501 mlip-studioThen open:
http://localhost:8501
For GPU-enabled Docker deployments, make sure the host has the NVIDIA Container Toolkit installed and adapt the image if a CUDA-specific PyTorch stack is required for your system.
Some upstream models may require accepting their original license terms or acceptable-use policies before use. Users are responsible for respecting the licenses of the underlying model families and datasets, including MACE, FAIRChem/UMA, ORB, MatterSim, SevenNet, PET/UPET, xTB, UFF, and related dependencies.
The public online instance is provided for free community access, but it is not intended for unlimited production workloads. For large systems, long trajectories, private data, or unrestricted benchmarking, run MLIP Studio locally.
MLIP Studio is released under the Apache Software License. See the ASL.md file for details.
Note that MLIP Studio provides access to several third-party model families, calculators, and datasets. These upstream components may be distributed under their own licenses or usage terms, which users are responsible for reviewing and following separately.
| Path | Description |
|---|---|
Home.py |
Main Streamlit application |
model_config.py |
Supported model definitions, model URLs/identifiers, citations, and sample structure list |
sample_structures/ |
Example molecules, crystals, surfaces, and interfaces |
requirements.txt |
Python dependencies |
Dockerfile |
Container build recipe |
mlip-studio-qm9-gap.pt |
In-house QM9 HOMO-LUMO gap model checkpoint |
If you use MLIP Studio in academic work, please cite the associated manuscript:
Manas Sharma, Sudeep Punnathanam, and Ananth Govind Rajan.
MLIP Studio: An Open Platform for Interactive Benchmarking and Simulations
Using Machine Learning Interatomic Potentials.
Citation details will be updated when a DOI or preprint is available.
MLIP Studio was developed by Dr. Manas Sharma at the Department of Chemical Engineering, Indian Institute of Science, Bengaluru, in the groups of Prof. Ananth Govind Rajan and Prof. Sudeep Punnathanam.