Source-linked AI summary
The Joint Automated Repository for Various Integrated Simulations (JARVIS) for data-driven materials design
Kamal Choudhary, Kevin F. Garrity, Andrew C. E. Reid, Brian DeCost, Adam J. Biacchi, Angela R. Hight Walker, Zachary Trautt, Jason Hattrick-Simpers, A. Gilad Kusne, Andrea Centrone, Albert Davydov, Jie Jiang, Ruth Pachter, Gowoon Cheon, Evan Reed, Ankit Agrawal, Xiaofeng Qian, Vinit Sharma, Houlong Zhuang, Sergei V. Kalinin, Bobby G. Sumpter, Ghanshyam Pilania, Pinar Acar, Subhasish Mandal, Kristjan Haule, David Vanderbilt, Karin Rabe, Francesca Tavazza
TL;DR
Materials design needs more integrated and comprehensive computational resources because existing DFT databases have important methodological and coverage limitations. JARVIS combines DFT, force-field, machine-learning, and workflow infrastructure, extending datasets and predictive models across specialized materials properties. The platform's distinctive scope includes beyond-standard electronic-structure data and ML models for rapid materials screening and property prediction.
Problem
Existing DFT databases have limitations from conventional DFT and omit factors such as van der Waals interactions, spin-orbit coupling, defects, temperature, and heterostructures.
Method
JARVIS integrates DFT, classical force-field, machine-learning, and automated computational tools into a materials-design infrastructure.
Results
JARVIS-DFT includes specialized datasets such as improved meta-GGA bandgaps, spin-orbit spillage, and frequency-dependent dielectric functions, while JARVIS-ML develops models for fast materials screening.
Takeaways & Limitations
The integrated datasets and models provide resources for materials screening, property prediction, and broader data-driven materials design.
Takeaways & Limitations
Experimental measurements or high-fidelity calculations for a given property are often scarce for evaluating computational predictions.
Abstract
from arXiv · showhide
The Joint Automated Repository for Various Integrated Simulations (JARVIS) is an integrated infrastructure to accelerate materials discovery and design using density functional theory (DFT), classical force-fields (FF), and machine learning (ML) techniques. JARVIS is motivated by the Materials Genome Initiative (MGI) principles of developing open-access databases and tools to reduce the cost and development time of materials discovery, optimization, and deployment. The major features of JARVIS are: JARVIS-DFT, JARVIS-FF, JARVIS-ML, and JARVIS-Tools. To date, JARVIS consists of 40,000 materials and 1 million calculated properties in JARVIS-DFT, 1,500 materials and 110 force-fields in JARVIS-FF, and 25 ML models for material-property predictions in JARVIS-ML, all of which are continuously expanding. JARVIS-Tools provides scripts and workflows for running and analyzing various simulations. We compare our computational data to experiments or high-fidelity computational methods wherever applicable to evaluate error/uncertainty in predictions. In addition to the existing workflows, the infrastructure can support a wide variety of other technologically important applications as part of the data-driven materials design paradigm. The JARVIS datasets and tools are publicly available at the website: https://jarvis.nist.gov .
Introduction
JARVIS addresses limitations in existing materials databases by integrating DFT, force-field, and machine-learning resources with automated tools. Its datasets extend beyond conventional properties to include specialized calculations, models, and workflows relevant to materials design.
- The Materials Genome Initiative seeks to accelerate materials discovery through computational, experimental, and data-analytics approaches.
- Existing DFT databases omit or simplify factors including van der Waals interactions, spin-orbit coupling, defects, temperature, low-dimensional thickness effects, and heterostructures.These omissions can be critical for realistic material applications.
- JARVIS integrates DFT, classical force-field, machine-learning, and workflow tools as an infrastructure for data-driven materials design.Its main components are JARVIS-DFT, JARVIS-FF, JARVIS-ML, and JARVIS-Tools.
- JARVIS-DFT uses van der Waals and beyond-GGA approaches and includes datasets for exfoliation energies, spin-orbit spillage, improved bandgaps, dielectric functions, SLME, infrared intensities, EFGs, heterojunctions, and Wannier Hamiltonians.These datasets are compared with experiments where possible to evaluate predictive accuracy.
- JARVIS-ML uses CFID descriptors to train models for rapid screening and energy-landscape mapping across properties including formation and exfoliation energies, bandgaps, magnetic moments, dielectric constants, and thermoelectric performance.The models also support interpretability analyses and STM-image analysis.
- JARVIS emphasizes integration and transparency through force-field resources, automated computational workflows, publicly disseminated datasets, and comparisons with experiments or high-fidelity calculations.The framework also identifies applications spanning experimental image interpretation and materials screening.
Overview of computational techniques
Materials simulation spans quantum, classical, mesoscale, finite-element, and engineering-design methods, each describing structure and properties differently. JARVIS focuses mainly on atomistic quantum/classical simulations and machine learning, with selected cross-method integrations.
- Range of computational techniques: Quantum, classical, mesoscale, finite-element, and engineering-design methods cover different length scales and use distinct descriptions of structure.For example, structure can mean electronic configuration, atomic arrangement, microstructure, phase-field segments, or mesh structure.
- Integration challenge: Realistic materials design requires integrating simulation methods while preserving the relevant physics across scales.Propagating results from one simulation into another is identified as a major multiscale-modeling challenge.
- JARVIS scope: Artificial-intelligence techniques can help integrate methods across simulation domains to a certain extent.The paper places machine learning within a broader effort to connect computational approaches.
- JARVIS scope: JARVIS primarily integrates atomistic classical and quantum simulations with machine learning.The infrastructure also attempts selected integrations with other simulation methods for specific applications.
- JARVIS scope: DFT elastic constants can feed orientation-distribution-function-based finite-element simulations.This is presented as an example of integrating atomistic data into another modeling approach.
Software and databases
JARVIS combines databases and computational tools for automating, analyzing, validating, and disseminating DFT, force-field, and machine-learning workflows. Its staged screening and cross-method comparisons support applications such as solar-cell materials discovery.
- Workflow: Screening proceeds from computationally inexpensive methods to more intensive methods, with experiments used where possible to assess accuracy and quality.Machine learning can be applied after generating a sufficiently large dataset.
- Derived applications: Solar-cell screening combines SLME, JARVIS-DFT dielectric functions and bandgaps, experimental comparison, beyond-DFT validation, and machine learning.Meta-GGA and GW calculations provide additional validation before ML accelerates future design.
- Software and databases: JARVIS integrates JARVIS-DFT, JARVIS-FF, JARVIS-ML, and JARVIS-Tools into a shared computational infrastructure.The repositories contain DFT data, force-field simulations, machine-learning models, and automation, post-processing, and dissemination tools.
- Integration: JARVIS-DFT data can feed JARVIS-FF and JARVIS-ML, while machine learning accelerates both DFT and force-field workflows.The stated data flow is one-directional from the lower-level DFT repository into FF and ML models.
- JARVIS-Tools: JARVIS-Tools automates DFT, molecular-dynamics, machine-learning, and Wannier calculations through predefined workflows and software interfaces.The framework supports packages including VASP, Quantum Espresso, LAMMPS, Scikit-learn, Keras, LightGBM, Wannier90, and Wanniertools.
- JARVIS-Tools: DFT workflows converge k-points and plane-wave cutoffs, relax energy, force, and stress, then compute properties on optimized structures.Subsequent calculations include band structures, dielectric functions, elastic and piezoelectric constants, and spin-orbit spillage.
JARVIS-DFT
JARVIS-DFT is a large repository extending conventional DFT databases across material classes, properties, functionals, and beyond-DFT calculations. It emphasizes van der Waals-aware calculations while documenting accuracy and computational limitations.
- Database scope: ≈ 40,000 materials and ≈ 1 million calculated properties are contained in JARVIS-DFT.The repository is mainly based on VASP calculations.
- DFT methodology: vdW-DF-OptB88 is used consistently across 3D, 2D, 1D, and 0D materials and is reported to predict accurate lattice parameters and energetics.The functional was selected for both van der Waals and non-van der Waals bonded materials.
- Database scope: JARVIS-DFT covers 3D bulk, 2D monolayer, 1D nanowire, and 0D molecular materials, although 3D and 2D data have primarily been distributed publicly.The database therefore spans multiple dimensionalities while public availability is uneven across them.
- Limitations: Bandgaps remain underestimated with vdW-DF-OptB88, while TBmBJ cannot describe excitonic electron-hole pairs in low-dimensional materials.Hybrid functionals and many-body methods are being generated as additional beyond-DFT datasets.
- Beyond-DFT methods: TBmBJ offers accuracy close to HSE06 at up to ten times lower computational expense for optical-gap prediction.The paper presents TBmBJ as a balance between computational cost and accuracy for high-throughput use.
- Properties and protocols: JARVIS-DFT adds properties and protocols including frequency-dependent dielectric functions, electric-field gradients, spin-orbit analyses, and automatic k-point convergence.These additions address capabilities beyond conventional database contents.
JARVIS-Beyond-DFT
JARVIS-Beyond-DFT addresses cases where semilocal DFT is insufficient, especially materials with relatively strong electron correlations. It organizes higher-level calculations for method selection and comparison with experiments.
- Method comparison: Semilocal DFT and DFT+U methods are described as fast and accurate for many structural parameters, whereas beyond-DFT methods extend benchmarking for particular cases.The paper distinguishes routine structural accuracy from cases requiring higher-level treatments.
- Motivation: Beyond-DFT methods may be needed for qualitative excited-state predictions when relatively strong electron correlations are present.The paper places this issue in systems such as superconductors, Mott insulators, heavy-fermion materials, semiconductors, photovoltaics, and topological Mott insulators.
- Database construction: JARVIS-Beyond-DFT builds a database of spectral functions and related quantities computed with meta-GGA, GW, hybrid functionals, and LDA+DMFT.These calculations support head-to-head comparisons on more than 100 materials.
- Research questions: The database investigates when beyond-DFT methods are necessary, which method to use, and how different methods compare with experiments.Target systems include transition-metal oxides, perovskites, nickelates, dichalcogenides, metals, and iron-based superconductors.
JARVIS-FF
JARVIS-FF is a collection of LAMMPS-based data and force-fields for large-scale atomistic materials simulations. It supports comparisons with DFT to assess force-field quality across multiple properties, while coverage remains application-dependent.
- Force-field performance can depend on the materials system, phenomena, application, and phase for which the force-field was designed.
- JARVIS-FF provides LAMMPS calculation-based data including crystal structures, formation energies, phonon densities of states, band structures, surface energies, and defect formation energies.
- ≈ 110 force-fields are included in the database, spanning EAM, MEAM, Bond-order and Tersoff, COMB, and ReaxFF types.
- JARVIS-FF converts crystal structures from JARVIS-DFT into LAMMPS inputs and runs calculations to produce material properties.
- Comparisons of force-field properties with corresponding DFT data help users analyze quality for applications involving convex hulls, elastic moduli, surface energies, and vacancy formation energies.
- JARVIS-FF is planned to include recently developed machine-learning force-fields.
JARVIS-ML
JARVIS-ML applies machine learning to discrete and image-based materials data using descriptors and trained models. Its reported strengths differ by task: regression performs well for several classical properties, while classification performs well for other properties and STM analysis.
- JARVIS-ML currently uses discrete targets from JARVIS-DFT for 3D and 2D materials and image-based scanning tunneling microscopy data.
- CFID descriptors encode chemical and structural information and can also be applied to molecules, proteins, point defects, free surfaces, and heterostructures.
- CFID provides 1557 descriptors per material, including chemical, simulation-box, radial-distribution, angle-distribution, and dihedral-angle descriptors.
- Trained models store parameters that rapidly predict properties of arbitrary compounds, and a web application hosts the models and trained-property list.
- Bulk modulus, maximum infrared active mode, and formation energies can be accurately trained, especially with regression models.
- Bandgaps, magnetic moments, piezoelectric coefficients, and thermoelectric coefficients achieve high-accuracy models for classification tasks only, while STM models classify image data.
Derived apps
Derived JARVIS applications extend the databases and tools into user-facing workflows for heterostructures, Wannier tight-binding calculations, orientation distributions, and simulation-quality assessment. These workflows connect material data to structural modeling, property prediction, and accuracy or precision analysis.
- JARVIS-Heterostructure characterizes heterojunction types and models interfaces for exfoliable 2D materials.
- The heterostructure app classifies systems as type-I, II, or III using Anderson’s rule and band alignment from JARVIS-DFT monolayers.
- JARVIS-WannierTB solves Wannier tight-binding Hamiltonians at arbitrary k-points for 3D and 2D materials and predicts band structures and densities of states on the fly.
- JARVIS-ODF is being developed to calculate volume-averaged meso-level properties, including elasto-plastic deformation, from single-crystal database data.
- Accuracy and precision analysis: Accuracy compares calculated values with experiments or high-fidelity theory, whereas precision concerns closeness among numerical approaches or simulation setups.
- Accuracy and precision analysis: Accuracy assessments are limited because high-quality experimental measurements or high-fidelity calculations are often available for only a few cases.
- Accuracy and precision analysis: FF simulation setups using the ‘refine’ and ‘box’ methods have minimal effects on FF-based predictions.
- Accuracy and precision analysis: 0.87+ precision is reported for all 2D Bravais lattices in STM classification, while regression-task precision analysis remains ongoing.
Future work
JARVIS provides publicly available databases, tools, documentation, and applications for materials design while continuing to expand its scope. Future work targets incomplete coverage, additional material properties, experiment integration, uncertainty analysis, and new ML capabilities.
- JARVIS databases are incomplete, creating an opportunity for substantial future expansion.
- Planned additions include defect and disorder properties, magnetic ordering, nonlinear optoelectronics, beyond-DFT calculations, temperature-dependent properties, experiment integration, and detailed uncertainty analysis.
- Future ML development will address data prediction, uncertainty quantification, explainable AI, and transfer-learning research.
- Website development includes on-the-fly calculation resources, advanced cross-database filtering, and visualization tools.
- JARVIS comprises JARVIS-DFT, JARVIS-FF, JARVIS-ML, and JARVIS-Tools, with generated data, notebooks, documentation, and calculation examples publicly available.
- The platform is intended to accelerate materials design and enhance industrial growth, with resources supporting participation by researchers worldwide.
- The authors state that publicly available data and resources will significantly accelerate future materials design in science and technology.
Methods
The study uses modular workflows spanning DFT, force-field, and related materials simulations, with explicit computational settings and scope constraints. These workflows calculate structural, electronic, thermoelectric, optoelectronic, defect, surface, phonon, and heterostructure properties.
- Workflow: The study manages, monitors, and analyzes calculations through a modular workflow made available in JARVIS-Tools.The workflow is also released through the JARVIS-Tools GitHub page.
- Scope and assumptions: The DFT calculations do not consider magnetic orderings beyond ferromagnetism, and nuclear spins are not explicitly included.The restriction on magnetic orderings is attributed to high computational cost.
- Property calculations: The workflows derive elastic, thermoelectric, optoelectronic, piezoelectric, dielectric, phonon, topological, exfoliation, heterostructure, and STM-related properties using specialized methods and software.Examples include finite differences, BoltzTraP with Constant Relaxation Time approximation, linear optics, DFPT, spin–orbit-coupling comparisons, and Wannier90.
- Scope and assumptions: Surface calculations include only unreconstructed surfaces without surface-segregation effects because the high-throughput approach does not yet account for element-dependent reconstructions.Surface energies are calculated from perfect bulk and surface structures for specific planes.
Machine learning training
Machine-learning workflows convert crystal structures into CFID descriptors and train predictive models on DFT-derived targets, with preprocessing and standard evaluation procedures. JARVIS-Tools supports these workflows alongside DFT and force-field calculations and related applications.
- Training infrastructure: Machine-learning models are mainly trained using Scikit-learn, Keras, and LightGBM, with trained models saved in pickle and joblib formats.Web applications are developed using JavaScript, Flask, and Django packages.
- Training data and descriptors: For scalar DFT properties such as formation energies, bandgaps, and exfoliation energies, crystal structures are converted into CFID input arrays and DFT data serve as targets.The data are split into training and test sets at a 90:10 ratio.
- Evaluation: Regression performance is generally reported with Mean Absolute Error or r2, while classification performance uses ROC Area Under Curve values between 0.5 and 1.0.Feature importance, k-fold cross-validation, and learning-curve analyses are also performed.
- Integrated applications: JARVIS primarily focuses on DFT and molecular-dynamics levels while integrating other simulation methods for specific applications.Several JARVIS components can work together to design optimized or completely new materials.
- Tools and applications: JARVIS-Tools provides workflows and examples for DFT, force-field, and machine-learning calculations, with databases and application snapshots organized across the infrastructure.The supplied passages reference workflows, databases, and derived applications rather than reporting a machine-learning benchmark result.