README.md

---
library_name: matgl
tags:
- matgl
- materials-science
- graph-neural-network
- machine-learning-interatomic-potential
- foundation-potential
- mlip
---
# TensorNet-PES-MatPES-PBE-2025.2-m

## Introduction

Pre-trained TensorNet foundation potential, i.e., universal machine learning interatomic potential trained on the MatPES-PBE-2025.2 dataset. This is a medium-size TensorNet variant (~1.07M parameters; `units=128, nblocks=3`), one block deeper than the standard `materialyze/TensorNet-PES-MatPES-PBE-2025.2` reference (0.84M).

## Potential

[matgl](https://github.com/materialyzeai/matgl) `Potential` model (version 3).

## Usage

```python
import matgl

model = matgl.load_model("materialyze/TensorNet-PES-MatPES-PBE-2025.2-m")
```

## Model Details

- Number of parameters: 1,067,906

## Metrics

| Split | Energy MAE (eV/atom) | Force MAE (eV/A) | Stress MAE (GPa) |
|---|---:|---:|---:|
| Train | 0.037036 | 0.111566 | 0.440905 |
| Validation | 0.037056 | 0.130899 | 0.592677 |
| Test | 0.037167 | 0.125499 | 0.592749 |

## Metadata

```json
{
  "dataset": "MatPES-PBE-2025.2",
}
```