Technology

Matlantis™: Core Technology and Mechanics

Our proprietary neural network potential (NNP)
significantly accelerates computation with high versatility and accuracy.

Matlantis supports large-scale materials discovery by simulating the behavior of various materials at an atomic level. The versatile atomistic simulator is powered by a neural network potential (NNP) in which a deep learning model is incorporated into a conventional atomic simulator.

Conventional physics-based atomistic simulation of material behavior requires high-cost calculations for solving the electronic state. Matlantis bypasses such calculations thanks to the deep learning model trained with energy, force, and other data necessary for simulating atomic behavior.

This allows researchers to quickly simulate a massive number of realistic, complex systems with relatively low calculation costs without compromising versatility.

There are some existing methods to perform atomic-scale simulation without the electronic state calculation. However, the governing equations are known to be too complex, and balancing versatility and accuracy is a long-standing challenge.

Matlantis can perform highly accurate simulations for atomic structures of any combination of 72 elements thanks to Preferred Networks (PFN)’s proprietary neural network architecture that has been trained with a vast amount of atomic structure data using PFN’s computer clusters.

PFN’s In-house Computer Cluster

Matlantis’s simulation results
accurately predicts DFT results

Research papers on Matlantis
Features

Matlantis: 3 Key Features

Matlantis supports companies exploring innovative materials for a sustainable future.

Versatility

image
Supports a wide range of elements and structures

Matlantis now supports 96 elements from the periodic table, covering all elements occurring in nature. This means users will encounter almost no restrictions regarding the types of elements they can work with. It can simulate the properties of any combination of atoms, including molecules and crystal systems, as well as unknown materials.

Speed

image
Over 10,000x faster than conventional methods

The atomistic simulation tasks that take hours to months using density functional theory (DFT) on a high-performance computer can be finished in only a few seconds using Matlantis.

User-Friendliness

image
Just open your browser to run simulations

Thanks to the pre-trained deep learning model, physical property calculation library, and high-performance computing environment, no hardware or software installations are required for performing simulations. Unlike conventional machine learning potentials, Matlantis requires no data collection or training by users.

Physical property calculation library

With additional codes, users can also simulate physical properties and phenomena from the NNP output (energy, force) without using the library.
Vibration analysis of organic molecules
Vibration mode・Vibration frequencyInfrared spectroscopy
Elastic constants
Elastic constantsYoung’s modulusBulk modulusPoisson’s ratio
Phonon dispersion and thermal properties of solid
Phonon dispersionHelmholtz free energySpecific heat
Dynamics calculation
Viscosity coefficientDiffusion coefficientSpecific heatThermal conductivity
Reaction pathway analysis
Nudged elastic band method (NEB)
How to

How to get started with Matlantis

STEP01

Log in to Matlantis
Log in to Matlantis capture

STEP02

Write your programs in Jupyter Notebook

Launch Jupyter Notebook and write simulation scripts in Python. Matlantis’s default environment allows you to run various packages such as first principles calculations and molecular dynamics methods using Python. Matlantis’ original physical property calculation library and sample notebooks are also available for you to immediately start your research.

Write your programs in Jupyter Notebook capture

STEP03

Execute the program

Matlantis executes the contents described in the Jupyter Notebook. It will take you only a few seconds to complete the structural optimization calculation.

Execute the program capture

STEP04

Export the result

Matlantis can also visualize the simulation results or perform further analysis on Jupyter Notebook.

Export the result capture

Try Matlantis

Contact Us

FAQ

Q.

Who is PFCC? What is your scope of work?

A.

PFCC, or Preferred Computational Chemistry, is the entity behind the Matlantis Service, established as a joint venture between Preferred Networks and ENEOS in 2021; we are dedicated to accelerating materials discovery for a sustainable future.
Category: #Company #Background

Q.

What is Matlantis, and what force field does it use?

A.

Matlantis is a high-speed universal atomistic simulator utilizing the machine-learned interatomic potential, Preferred Potential (PFP), capable of handling up to 72 elements.
Category: #ForceField #Product #Service

Q.

What data preparation is necessary to utilize Matlantis to simulate proprietary materials in development?

A.

Matlantis offers pre-trained universal potentials, eliminating the need for our users to create and train their datasets.
Category: #ForceField #Data

Q.

How does Matlantis handle the selection of interatomic potentials, and how many different potentials are available?

A.

Matlantis offers a universal potential (PFP) suitable for various systems as well as specific calculation modes to fine-tune calculation conditions and, as such, boost simulation accuracy. For details of calculation modes, see below:

– MOLECULE: molecule (non-periodic) systems
– CRYSTAL: crystal (periodic) systems with the Hubbard U correction
– CRYSTAL_PLUS_D3: crystal (periodic) systems with the Hubbard U correction + DFT-D3 dispersion correction
– CRYSTAL_U0: crystal (periodic) systems without the Hubbard U correction
– CRYSTAL_U0_PLUS_D3: crystal (periodic) systems without the Hubbard U correction + DFT-D3 dispersion correction

Category: #ForceField #Accuracy #Capacity

Q.

What functionals do you use for DFT calculations on training data?

A.

For periodic systems, we adopt the PBE functional with PAW pseudopotentials. For non-periodic systems, the data are constructed by DFT calculations using the ωB97xd functional and the 6-31G(d) basis set.
Category: #ForceField #Data #Modeling

Q.

What level of accuracy can I expect from Matlantis calculations? And where can users find validation results for Preferred Potential (PFP)?

A.

Matlantis achieves an approximate mean absolute error (MAE) of 0.03 eV/atom for random data and 0.01 eV/atom for bulk data. For validation details, refer to our recent benchmarking results: PFP v5 Validation; previous reports are available here: PFP v4 Validation
Category: #ForceField #Accuracy

Q.

What is the maximum number of atoms Matlantis can handle in calculations?

A.

Matlantis can typically handle several thousand to 10,000 atoms, depending on the calculation requirements. For example, we can calculate up to 16,000 atoms in the case of SiO₂ crystals.
Category: #ForceField #Capacity #System

Q.

Can Matlantis be applied to resin and polymer research?

A.

Matlantis can simulate monomer reactivity or polymer substructures, prioritizing accuracy and universality. In addition, users can apply Matlantis to systems containing both organic and inorganic atoms.
Category: #Application

Q.

What are the hardware requirements for using Matlantis?

A.

Matlantis simulation is executed in the cloud, users do not need demanding hardware to access your tenant. It is accessible via standard terminal (such as a laptop) and network environments, with a Google Account or Azure AD account authentication.
Category: #Accessibility #Authentication

Q.

How many GPUs are available for use with Matlantis? How can GPU acceleration be utilized for simulation jobs in Matlantis?

A.

GPU availability depends on the chosen plans. Each plan offers sufficient calculation resources to meet our customers’ varied needs. Jobs are submitted to the Matlantis batch inference system for efficient GPU acceleration.
Category: #ComputingResources #Hardware #GPU

Q.

Tell me more about Matlantis’ training datasets.

A.

To train our neural network potential PreFerred Potential (PFP), we utilized our in-house computing resources to collect a vast amount of data. These datasets are proprietary, and for more details, please refer to the supplementary materials provided in our publication: Towards universal neural network potential for material discovery applicable to arbitrary combination of 45 elements | Nature Communications
Category: #ForceField #Datasets #Modeling

Q.

How does Matlantis address Van der Waals interactions?

A.

Matlantis offers Grimme’s DFT-D3 energy correction to handle dispersion interactions such as Van der Waals forces.
Category: #ForceField #Features #Capability

Q.

Is the Hubbard U+ correction feature available in Matlantis?

A.

Yes, it is. Matlantis offers Hubbard correction calculation mode to mitigate some of the deficiencies of local and semi-local exchange-correlation functionals while maintaining computational efficiency.
Category: #ForceField #Features #Capability

Q.

What is the current cutoff distance used in Matlantis calculations?

A.

Since the release of PFP v6.0.0, we have upgraded the maximum cutoff distance to nine Angstroms; however, the cutoff is adjusted depending on the system’s density.
Category: #ForceField

Q.

What simulation engines are available in Matlantis?

A.

ASE and LAMMPS(*) are available; we provide detailed tutorials/instructions on implementing them in Matlantis.
*provided as an experimental feature.
Category: #LAMMPS #ASE #Compatibility #Software

Q.

Does Matlantis support DFT calculations?

A.

No, Matlantis does not support DFT calculations at the moment.
Category: #DFT #Compatibility #Software

Q.

Which cloud provider does Matlantis utilize? How secure is user data in the Matlantis cloud?

A.

Matlantis utilizes AWS, GCP, and PFN’s clusters for cloud services, offering secure data storage comparable to or exceeding on-premise solutions. All network communications in the service and stored data are encrypted.
Category: #Security #System #Hardware #Cloud #OnPremise

Q.

How many computational resources does Matlantis possess?

A.

We cannot publicly disclose the instance name, but we are happy to schedule a meeting to reveal more specs of the computational resources.
Category: #System #Hardware

Q.

How can I obtain pricing and licensing information for Matlantis? Is a pay-as-you-go option available?

A.

Please contact our sales team to schedule an appointment to discuss our price plans.
Category: #System #Pricing

Q.

Are there tutorials or documentation available for Matlantis reference?

A.

Yes, the Matlantis tutorial is accessible through the provided link: Atomistic Simulation Tutorial
Category: #System #Tutorial #Support

Q.

Where can I find past webinars or seminars hosted by PFCC or Matlantis users?

A.

We archived our previous webinars on the Matlantis YouTube account and Matlantis (@matlantis)on Speaker Deck.
Category: #System #Events #Webinars #Seminars

Q.

What does the Matlantis GUI look like?

A.

Matlantis offers Jupyter Lab for Python coding and simulation setup.
Category: #System #Software #GUI

Q.

Where can I find publications or research cases related to Matlantis?

A.

Publications can be found on the Matlantis publication page, while research cases are available on the Matlantis case studies page.
Category: #System #Publications #CaseStudies

Q.

Can individuals from specific countries or regions access Matlantis?

A.

For the time being, Matlantis is available in North America, Europe, and some APAC regions. Please contact our sales team for inquiries regarding Matlantis’ availability in specific areas.
Category: #System #Avalability