For materials science, you can choose between generative AI research models or dedicated enterprise platforms. Microsoft's MatterGen and Google DeepMind's GNoME are top choices for generative design and stability testing. If you need a production-ready system for industrial formulation, material design, or property prediction, platforms like Citrine Informatics, Schrödinger, and Polymerize offer integrated tools for your specific R&D workflows.
Brands AI recommends here
Mentioned in
1
Schrödinger
Best for high-precision molecular simulation combined with machine learning. It is ideal for atomic-level materials modeling and physics-based design, though it serves more specialized research needs.
46%
2Citrine InformaticsThe primary choice for industrial R&D teams needing enterprise-ready materials informatics. It excels at data-driven property prediction, experimental design, and integrating ML into existing industrial workflows.37%
Using AI for materials science has evolved rapidly from simple database screening to sophisticated atomic foundation models and generative inverse-design frameworks . Several major tech enterprises, specialized startups, and open-source ecosystems now offer pre-trained models capable of predicting chemical properties, simulating atomic forces, or generating novel crystalline and molecular structures.
The leading providers and their flagship models mapped to chemical and material properties include:
Microsoft Research provides Microsoft MatterGen and Microsoft MatterSim. is a generative diffusion model designed for inverse materials design, allowing you to directly generate stable inorganic crystal structures steered toward specific chemical, electronic, magnetic, or mechanical constraints. functions as a deep-learning atomistic simulator capable of predicting energies, forces, and dynamic behavior over wide ranges of temperature and pressure.
Orbital Materials offers Orb , an open-source universal interatomic potential model designed to simulate advanced materials (such as catalysts, semiconductors, and energy-storage mediums). Orb is optimized for fast and accurate large-scale molecular dynamics and energy estimations, showing competitive performance on benchmarks like Matbench Discovery. You can explore the project via Orbital Industries.
Google DeepMind developed GNoME (Graph Networks for Materials Exploration) , which famously expanded humanity's known stable materials catalog by hundreds of thousands of crystal structures. While parts of its ecosystem tie closely into Google's internal infrastructure, their foundational approaches set the standard for predicting formation energies and chemical stability using graph neural networks.
Entalpic specializes in an AI platform tailored for surface-driven industrial processes and catalysis. Their engine combines quantum chemistry simulations with multimodal datasets (patents, literature, and quantum data) to screen and generate viable chemical and material candidates under real industrial constraints. More details are available on Entalpic.
Dassault Systèmes (BIOVIA) integrates Scientific AI into its enterprise modeling suites. Unlike general-purpose language models, BIOVIA's tools use domain-specific training to assist in molecular generation, formulation design, and predicting chemical behaviors within commercial R&D pipelines. Discover their approach via Dassault Systèmes BIOVIA.
To help narrow down the right tool for your project, tell me:
Are you focusing on crystalline solids / inorganic materials or organic molecules / polymers?
Do you need a model for property prediction/simulation or generative inverse design?
If you are planning to use AI for materials science, there are several classes of models depending on whether you want property prediction, molecular/material generation, simulation acceleration, or literature/knowledge mining. The strongest options today are usually trained on combinations of DFT calculations, crystal structures, molecular databases, and experimentally measured properties.
1. Materials foundation models (crystals, inorganic materials)
Provider / model
What it does
Typical properties/tasks
Microsoft MatterGen
Generates new inorganic crystal structures conditioned on target properties
Band gap, bulk modulus, magnetic properties, chemical systems, crystal structures
Google DeepMind GNoME
Large-scale discovery of stable crystal candidates
MatterGen is particularly relevant if your goal is inverse materials design—for example: “find a crystal with a band gap near X, high bulk modulus, or a specific chemistry.” It provides pretrained checkpoints conditioned on properties such as band gap, magnetic density, bulk modulus, chemical system, and energy above hull.
Transformer trained on SMILES strings; molecular property prediction
Uni-Mol
3D molecular representation learning
These are more appropriate for:
polymers
organic semiconductors
catalysts
solvents
drug/material interfaces
molecular property prediction
3. Commercial AI platforms for materials discovery
Provider
Offering
Schrödinger
Computational chemistry and materials simulation workflows
Citrine Informatics
Materials informatics platform for property prediction and experimental optimization
Materials Design
Atomistic simulation software and AI-assisted materials modeling
IBM Research
4. Open datasets commonly used for training/fine-tuning
Materials Project — crystal structures and DFT-derived properties.
Open Quantum Materials Database — computational materials properties.
Alexandria Materials Database — DFT-relaxed structures used in some newer materials models.
PubChem — chemical structures and molecular properties.
QM9 — quantum chemistry benchmark dataset.
How I would choose a model
“Predict properties of my existing compounds”
Start with Chemprop, M3GNet, MACE, or similar graph models.
“Design new battery materials / catalysts / superconductors”
Look at MatterGen, GNoME-derived workflows, M3GNet, and DFT validation loops.
“Optimize polymers or organic materials”
Use molecular foundation models (MolFormer, Uni-Mol, graph neural networks) plus your experimental data.
“Run molecular dynamics faster than DFT”
Use ML interatomic potentials (MACE, M3GNet, MatterSim).
“Build an internal AI materials lab”
A common architecture is:
Foundation model → generate candidates
ML surrogate → rank candidates
DFT/experiments → validate
Active learning → improve model
If you tell me your target area (battery materials, polymers, semiconductors, catalysts, metals, ceramics, etc.) and whether you want commercial APIs or open-source models, I can narrow this to the most suitable options.
Several leading tech enterprises, research labs, and informatics platforms offer advanced AI models specifically trained on chemical structures, crystal symmetries, and physical properties for materials science.
These platforms bypass standard generic LLMs by using graph neural networks (GNNs) and atomistic foundation models trained on quantum mechanics and Density Functional Theory (DFT) datasets.
Generative models (MatterGen ) and atomistic simulation engines (MatterSim ) conditioned on chemistry and property targets.
De novo generation of custom materials with target constraints.
Architectural Types of Chemical AI Models
Property Predictors: Models like ALIGNN and SchNet take atomic coordinates and map them directly to electrical, optical, or thermal properties.
Generative & Inverse Design Models: Tools like MatterGen accept target criteria (such as specific elasticity or conductivity) and generate new atomic configurations matching that chemical space.
Chemical Language Models: Models like polyBERT treat polymer strings or SMILES notations as a specialized language to infer behavior without full physical simulations.
If you can share what specific properties (e.g., thermal, mechanical, optical) and what class of material (e.g., polymers, inorganic crystals, alloys) you are targeting, I can recommend the most precise open-source or commercial tool for your workflow.
Yes. There is now a fairly broad ecosystem of AI models specifically trained on chemical structures, crystal structures, and/or calculated materials properties. The right choice depends heavily on whether you want property prediction, materials generation, or simulation.
Provider / model
What it is useful for
Example properties / capabilities
Access
Microsoft — MatterGen
Generating new inorganic materials subject to constraints
Band gap, bulk modulus, magnetic density, chemical system, formation energy
Open model + Azure
Microsoft — MatterSim
Predicting/simulating materials behavior
Atomic interactions across elements, temperatures and pressures; materials-property prediction
Formation energy, band gap, stability, elasticity, etc.
Open data
Specialized materials/chemistry ML models
Property prediction from molecular/crystal structures
Varies: solubility, electronic properties, mechanical properties, reaction properties, etc.
Often open-source/research
The one I'd investigate first: MatterGen
Microsoft's MatterGen is particularly interesting if by "use AI for material science" you mean designing new materials rather than simply predicting properties. It is a diffusion model trained for inorganic materials and can be conditioned on chemical system and properties such as band gap, bulk modulus, magnetic density, and energy above the convex hull.
Microsoft provides pretrained checkpoints and, importantly, the model can be fine-tuned on your own property-labeled data.
MatterSim is more relevant. It is designed as a deep-learning atomistic model for materials simulation and property prediction across a wide range of elements, temperatures, and pressures. Microsoft reports applications spanning metals, oxides, sulfides, halides, crystals, amorphous solids, and liquids.
Google DeepMind / GNoME
GNoME is another major example. DeepMind trained it on crystal structures and stability data and used an active-learning loop involving DFT calculations to discover and evaluate candidate stable crystals.
The distinction is useful:
MatterGen → "Generate a material with these properties."
MatterSim → "Predict/simulate properties and atomic interactions."
GNoME → "Find and screen potentially stable crystal structures."
If you're starting a materials-AI research program, I'd actually build the stack around Materials Project data + an atomistic/property model + a generative model, rather than betting everything on a general-purpose LLM. MatterGen is especially attractive because its code supports fine-tuning on your own property data.
If you tell me whether you're working on batteries, catalysts, polymers, semiconductors, alloys, ceramics, or something else, I can give you a more specific shortlist of the 5–10 strongest models/datasets/APIs for that area, including which are open-weight versus commercial.
A robust ecosystem of public institutions, tech giants, and specialized informatics firms offer AI models and pre-trained datasets specifically mapped to chemical compositions, crystal structures, and physical properties.
The primary providers and platforms offering AI models trained on chemical and material properties include:
Google DeepMind (GNoME & Materials Discovery)
What they offer: Google DeepMind released GNoME (Graph Networks for Exploring Materials) , a deep learning tool that dramatically scaled the discovery of stable inorganic crystals. They have predicted millions of new structures and identified hundreds of thousands of stable materials primed for applications like batteries, superconductors, and supercomputers.
What they offer: Run by Lawrence Berkeley National Laboratory and international collaborators, the Materials Project hosts computed properties (thermodynamic stability, electronic band structures, elastic tensors) for hundreds of thousands of inorganic materials and molecules. They provide standardized, "AI-ready" datasets and foundational machine-learning interatomic potentials (MLIPs) specifically designed to map chemical composition to physical behavior.
What they offer: Microsoft integrates AI with high-performance quantum chemistry simulations. Azure Quantum Elements combines generative AI with density functional theory (DFT) to accelerate the screening of chemical compounds, acting as a copilot for chemical and material design.
What they offer: A commercial platform built specifically for materials and chemicals development. The Citrine Platform uses proprietary machine learning models trained on vast proprietary and public materials data to predict mechanical, chemical, and thermal properties, helping R&D teams optimize alloy, polymer, and chemical formulations.
What they offer: The academic and open-source community maintains specialized graph neural networks (GNNs) pre-trained on chemical properties. Examples include MEGNet and CGCNN (for crystal graph convolutions) and ALIGNN (interatomic line graph networks for predicting alloy and ionic properties).
Access: These models are typically open-source and hosted on GitHub or via materials science Python libraries like matminer or pymatgen.
If you'd like to narrow this down, let me know:
What specific material class you are targeting (e.g., polymers, crystalline alloys, battery electrolytes)?
Are you looking for an open-source model to run locally or a commercial platform?
Yes. There is now a fairly substantial ecosystem of AI models trained specifically on chemical composition, crystal structures, atomic interactions, and computed materials properties. The right choice depends heavily on whether you want prediction, simulation, or generation.
Provider / model
What it is useful for
Property / chemistry focus
Access
Microsoft — MatterSim
Atomistic simulation and property prediction
Energies, forces, structures, behavior across elements, temperature and pressure
Open models
Microsoft — MatterGen
Generating new inorganic materials
Chemical system, band gap, magnetic density, bulk modulus, stability-related properties
Open model + Azure
Materials Project
Materials-property data rather than one model
Formation energy, band gaps, stability, crystal structures, etc.
If you're building an AI-for-materials research stack, I'd look closely at Microsoft's two models.
MatterSim is an atomistic foundation model intended for simulation and property prediction. It covers a broad portion of the periodic table and conditions up to 5,000 K and 1,000 GPa, and can be customized with domain-specific data.
MatterGen goes in the opposite direction: rather than predicting a property of a material you give it, it can generate candidate inorganic materials subject to desired properties or chemistry. Released checkpoints include conditioning on chemical system, DFT band gap, magnetic density, bulk modulus, and energy above the convex hull.
One important limitation is that MatterGen is aimed at inorganic crystalline materials; its current documented scope excludes organic crystals and non-crystalline materials and is limited to structures with up to 20 atoms in the unit cell.
If your goal is chemistry/materials research
I'd divide the landscape into four categories:
Property prediction: Given a molecule/material, predict band gap, formation energy, stability, mechanical properties, etc.
Atomistic simulation: Learn energies/forces so you can perform much cheaper molecular dynamics or structure relaxation — MatterSim is particularly relevant here.
Inverse materials design: Specify desired properties and generate candidate structures — MatterGen is a prominent example.
Materials knowledge/retrieval: Combine an LLM with databases such as the Materials Project to answer questions about known compounds and properties.
The Open Catalyst Project is another useful ecosystem if your work involves catalysis. Its OC20/OC22 datasets contain roughly 1.3 million molecular relaxations and more than 260 million DFT calculations, specifically aimed at ML approaches to catalyst simulation.
If you tell me what kind of materials you're working with (e.g. batteries, polymers, semiconductors, catalysts, alloys, ceramics, organic molecules), I can narrow this down to the most relevant models and give you a practical 2026 AI-materials stack—including datasets, pretrained models, GPUs, and which models can actually be fine-tuned on your own experimental data.
Yes. There is now a fairly strong ecosystem of AI models trained on chemical/materials data, but they fall into different categories: property prediction, atomistic simulation, materials generation, and literature/knowledge extraction.
Strong options to investigate
Microsoft — MatterGen: Probably the most directly relevant if your goal is inverse materials design. It is a diffusion model for inorganic crystals and can generate structures conditioned on properties such as band gap, bulk modulus, magnetic density, chemical system, and energy above hull. It is open-source and can be fine-tuned on your own property data.
Microsoft — MatterSim: Better suited to predicting/characterizing material properties and atomic interactions rather than generating new compounds. Microsoft describes it as a deep-learning atomistic model spanning many elements, temperatures and pressures, including crystals, amorphous materials, liquids, metals, oxides, sulfides and halides.
IBM — Foundation Models for Materials (FM4M): A multimodal family of models covering representations including SMILES, SELFIES, molecular graphs, 3-D atomic positions and 3-D density grids. The models are intended to be fine-tuned for predictive and generative chemistry/materials tasks.
MatSciBERT: A materials-science language model rather than an atomistic property model. It was pretrained on a materials-science corpus and is useful for extracting chemical/materials information from papers, classification, entity recognition and knowledge-graph construction.
A useful way to think about the landscape
If you want to…
Look at
Predict material properties
MatterSim, other atomistic foundation models
Generate new inorganic crystals
MatterGen
Generate materials with target properties
MatterGen
Work with molecular/chemical representations
IBM FM4M
Mine scientific literature
MatSciBERT
One particularly interesting development is that property-conditioned generation is becoming practical. MatterGen already has pretrained/fine-tuned checkpoints for things like band gap, bulk modulus and magnetic density, and its architecture allows you to fine-tune against your own labeled property.
If you're planning an actual materials-AI research stack, I would distinguish between three questions:
What material should I make? → generative model such as MatterGen
What properties will it have? → atomistic/property foundation model such as MatterSim
What does the literature already know about it? → materials/chemistry language models
There is also a rapidly developing class of foundation models specifically for materials property prediction with small experimental datasets, which may be especially relevant if you have only hundreds or thousands of measurements rather than millions. A 2026 study specifically investigates in-context learning for this setting.
If you tell me what kind of materials you're working with (e.g. batteries, polymers, catalysts, semiconductors, alloys, ceramics) and whether you want prediction vs. generation, I can narrow this to the 5–10 best models/datasets/APIs for your use case, including which are open-source and which are commercial.
Several major tech institutions, research labs, and specialized informatics companies offer advanced AI models trained on chemical compositions, crystal structures, and physical-chemical properties.
Leading providers of these capabilities include:
Google DeepMind (GNoME & MEGNet / MatGL) : DeepMind developed GNoME (Graph Networks for Materials Exploration), a deep learning tool that dramatically expanded humanity's catalog of stable crystals. They also support MatGL (Materials Graph Library), which implements graph neural networks like M3GNet and CHGNet to predict properties, atomic forces, and trajectories across the periodic table.
Microsoft Research (MatterGen & MatterSim) : Microsoft's AI for Science initiative features MatterGen , a generative diffusion model that designs entirely new materials from scratch based on targeted chemical, mechanical, electronic, or magnetic constraints. It pairs with MatterSim , an atomistic foundation model that predicts stability and behavior using quantum-level accuracy.
Citrine Informatics : A commercial software platform built specifically for chemistry and materials science. The Citrine Platform ingests proprietary or public chemical/material data and builds customized machine learning models with uncertainty quantification to predict physical properties, optimize chemical formulations, and guide synthesis.
The Materials Project & Open Catalyst Project : Run collaboratively (with major contributions from national labs and universities), these platforms host open-access computed databases and pre-trained machine learning models (like MEGNet and various graph neural networks) specifically mapped to chemical formulas, crystal symmetry, and formation energies.
To help narrow down the best tool for you, could you share:
Are you looking for an open-source model to run locally, or a commercial platform?
What specific properties (e.g., thermal, electronic, mechanical) or material classes (polymers, inorganic crystals, catalysts) are you targeting?
Yes. There is now a fairly broad ecosystem of AI models pretrained on chemical/molecular or materials data, but they fall into different categories. The right choice depends heavily on whether you're working with small molecules, polymers, inorganic crystals, batteries, catalysts, or formulations.
Strong options to investigate
Provider / model
Best for
What it's trained on / does
Microsoft — MatterGen
Inorganic materials discovery
Generative model for crystal structures; can condition generation on properties such as band gap, bulk modulus, magnetic density, chemical system, and energy above hull.
Microsoft — MatterSim / MatterSim-MT
Predicting material properties
Foundation model aimed at atomistic simulation and multi-task materials characterization; particularly interesting if your goal is property prediction rather than generation.
IBM — FM4M / SMI-TED
Molecular & chemical-property prediction
SMI-TED was pretrained on ~91M SMILES samples from PubChem and supports tasks including quantum-property prediction.
ChemFM
General molecular chemistry
3B-parameter chemical foundation model pretrained on 178M molecules; designed for downstream property prediction, molecular generation and reaction prediction.
Schrödinger — Formulation ML
Formulations / mixtures
Commercial ML specifically for relating ingredient structures/compositions to formulation properties and screening new formulations.
POLYT5
Polymers
Chemical foundation model specifically trained to understand and generate polymer structures.
If you're building an AI materials-science stack
I'd divide the landscape into four layers:
1. Molecular models — SMILES/SELFIES-based models such as IBM's SMI-TED and ChemFM.
Good for predicting things like molecular properties, reactions, quantum properties, toxicity, solubility, etc.
2. Atomistic/materials models — MatterSim and related machine-learned interatomic potentials.
These are much more relevant when you care about 3D structures, energies, forces, phonons, stability, and physical properties. A 2026 review describes this rapidly developing area as foundation models for atomistic simulation of chemistry and materials.
3. Generative materials models — MatterGen is a particularly notable example. Rather than asking "what property does this material have?", you can ask something closer to "generate a material with this target property."
4. Specialized commercial platforms — Schrödinger and chemistry-data platforms such as Elsevier's Reaxys provide more complete R&D workflows rather than just a downloadable model.
One important distinction: "trained on chemical properties" doesn't necessarily mean the model was directly trained to predict those properties. Some models learn chemical representations from huge molecular datasets and are subsequently fine-tuned on property datasets; others are trained directly on DFT energies/forces/properties.
If you tell me what kind of material you're working on (e.g. batteries, polymers, catalysts, semiconductors, alloys, pharmaceuticals) and whether you want property prediction, inverse design/generation, simulation, or an API/commercial product, I can narrow this down to the 5–10 models/platforms I'd actually evaluate, including training data, licenses, APIs, compute requirements, and benchmark performance.
Using AI for materials science and chemical property prediction involves a mix of enterprise software platforms, open research datasets, and cutting-edge deep learning models trained on crystal structures, thermodynamics, and quantum chemistry.
Major organizations, platforms, and research groups offer AI models and infrastructure specifically trained or customizable for chemical and material properties:
1. Enterprise Materials Informatics Platforms
Citrine Informatics : A leading enterprise software platform built specifically for chemistry and materials development. It allows R&D teams to ingest proprietary lab data, train machine learning models on specific target chemical/physical properties (e.g., tensile strength, conductivity, thermal stability), and optimize formulations using AI via their Citrine VirtualLab and Citrine DataManager tools.
Schrödinger : Known heavily in computational chemistry and drug/material discovery, Schrödinger combines physics-based digital molecular modeling with machine learning models. Their platform predicts chemical properties, binding affinities, and material behaviors by tightly coupling quantum mechanics (QM) with machine learning layers.
Materials Zone : Offers an AI-driven data and collaboration platform tailored for materials science, helping teams centralize characterization data and build predictive models for material performance and manufacturing outcomes.
2. Tech Giants & Open-Source AI Research Labs
Google DeepMind (GNoME & Materials Discovery) : DeepMind released GNoME (Graph Networks for Exploration of Materials) , a deep learning tool trained extensively on crystal structures and thermodynamic stability. They open-sourced data for hundreds of thousands of stable inorganic materials to help predict new crystal phases and chemical compositions. Their work heavily integrates with the Materials Project.
Entalpic : An emerging AI platform focused on accelerating chemistry and materials R&D. It combines generative models and physics-based simulations with atomistic machine learning models via an "Atomistic Discovery Engine" to map chemical space and predict candidate material behaviors.
3. Open-Source Libraries & Foundation Models for Atoms
Open-Source Universal Interatomic Potentials (e.g., M3GNet, CHGNet, NequIP) : Developed by various academic and research groups (such as UC Berkeley and Lawrence Berkeley National Lab), these graph neural networks (GNNs) are trained on massive density functional theory (DFT) databases like the Materials Project. They can instantly predict forces, energies, and structural relaxations for arbitrary chemical compositions without running slow, expensive quantum-mechanical simulations from scratch.
PyMatGen (Python Materials Genomics) : The industry-standard open-source Python library used to interface with crystal structures, manipulate molecular/atomic coordinates, and feed data into machine learning pipelines for property forecasting.
If you'd like to narrow this down, let me know:
Are you looking to predict properties for small molecules/polymers or inorganic crystals/metals?
Do you need an off-the-shelf API/open-source model or an enterprise platform for proprietary lab data?
