The matterverse is vast and complex. It comprises the infinite combinations of elements of the periodic table in ordered and disordered arrangements. While ab initio techniques can accurately probe the matterverse, the scope of our exploration has been bottlenecked by their high cost and poor scaling. is a database of yet-to-be-sythesized materials predicted using state-of-the-art machine learning (ML) algorithms. At the present moment, It aims to revolutionize our ability to explore the matterverse at unprecedented scales and accuracy. Today, graph deep learning models today can achieve predictive accuracies that allow us to search vast chemical spaces for novel technological materials.

Database statistics

No. of materials


No. of elements


No. of chemistries

> 40,000

Machine Learning Models

One of the main underlying technologies of is M3GNet. M3GNet is a cutting-edge materials graph neural network architecture that incorporates 3-body interactions developed by the Materials Virtual Lab.

One of the key advantages of M3GNet is its flexibility and scalability. For instance, we have trained a M3GNet universal interatomic potential (IAP) for 89 elements of the periodic table using energy, force and stress data from density functional theory (DFT) relaxations performed by the Materials Project (MP) since 2012. IAPs, which describe the potential energy surface of a collection of atoms, are a fundamental input for atomistic simulations. The M3GNet universal IAP has broad applications in structural relaxation, dynamic simulations and property prediction of materials across diverse chemical spaces. For example, it can be used to generate equilibrium structures for any arbitrary crystal structure without expensive ab initio calculations, providing a pathway to massive-scale materials design.

Beyond IAPs, the M3GNet architecture can also be used to develop property prediction models. The table below summarizes the performance of M3GNet models used to provide predictions for a common set of properties for all materials in

Property Dataset size MAE/AUC
MP Formation Energy Ef (meV/atom) 132,752 19.5±0.2
MP Band Gap Eg (meV/atom) 106,113 0.183±0.005
MP Bulk Modulus, log(KVRH) (GPa) 10,987 0.058±0.003
MP Shear Modulus, log(GVRH) (GPa) 10,987 0.086±0.002
MP Metal Prob. 106,113 0.958±0.001

For more information about M3GNet, please refer the following published work:

Chen, C.; Ong, S. P. A Universal Graph Deep Learning Interatomic Potential for the Periodic Table. arXiv:2202.02450 [cond-mat, physics:physics] 2022.

How can I reproduce your results?

The M3GNet source code is open-sourced on Github and the package can be installed via "pip install m3gnet". An example code to reproduce a relaxation for one of the materials in the matterverse database is given below.

# We first get the parent structure from MP.
# Note that MP structures do not contain oxidation states, unlike the original ICSD source.
from pymatgen.ext.matproj import MPRester
mpr = MPRester()
mpr.query({"icsd_ids": 1570}, ["material_id"])
r = mpr.query({"icsd_ids": 1570}, ["material_id"])
mpid = r[0]["material_id"]
parent = mpr.get_structure_by_material_id(mpid)

# Do the substitution as shown in the page.
parent.replace_species({'K': 'Li', 'O': 'O', 'Tl': 'Li'})

# Relaxation using M3GNet
from m3gnet.models import Relaxer
relaxer = Relaxer()
results = relaxer.relax(parent)