I am planning to use AI for material science. Who offers models trained on chemical properties?

Yes. If by “models trained on chemical properties” you mean models that understand composition, crystal structure, energetics, and experimentally/DFT-derived material properties, there are several serious options.

The two I'd investigate first

1. MatterGen — probably the closest match if your goal is “tell the AI what properties I want, and have it propose candidate materials.” Microsoft says it can condition generation on chemistry, symmetry, and mechanical, electronic, and magnetic properties. Its pretrained checkpoints are publicly available, and the implementation is MIT licensed.

2. MatterSim — better if you already have candidate materials and want “predict their properties / simulate their behavior.” It covers a broad range of elements and conditions, including 0–5,000 K and pressures up to 1,000 GPa, and can be fine-tuned with domain-specific data.

A useful distinction is:

LLM for chemistry → understands chemical literature, synthesis procedures, papers, etc.
Materials foundation model → understands atoms + composition + structure + physical properties.
Generative materials model → can propose entirely new structures satisfying property constraints.

For actual materials science, I'd generally start with the second/third category rather than a general chemistry LLM.

If you tell me whether you're working on batteries, catalysts, polymers, semiconductors, alloys, ceramics, or something else, I can give you a more targeted list of the strongest models and commercial APIs available in 2026.

A robust ecosystem of tech giants, specialized startups, national laboratories, and academic platforms offer advanced AI models trained on chemical compositions, crystal structures, and physicochemical properties.

The primary providers of AI models and platforms designed for chemical and materials property prediction include:

• CuspAI Focuses on AI-driven materials discovery, building an extensive partner network (Materials Foundry) to predict and experimentally validate chemical behaviors and material properties.
• Google DeepMind Developed GNoME (Graph Networks for Materials Exploration), an open-type AI model trained to predict the stability and structural chemical properties of inorganic compounds, scaling down millions of candidate structures.
• Microsoft Research Offers MatterGen (a generative AI tool for inverse-design of materials based on target chemical, mechanical, or electronic constraints) and MatterSim for atomistic property modeling.
• NVIDIA Provides the ALCHEMI Toolkit , a PyTorch-native, GPU-accelerated framework for Machine Learning Interatomic Potentials (MLIP) that bridges natural-language prompting and materials simulations.
• U.S. Department of Energy (DOE) / ALCF Released open-source foundation model suites like MIST via national supercomputing resources, designed explicitly for molecular property prediction and chemical system screening.
• Academic & Open-Source Architectures Frameworks utilizing graph neural networks (GNNs) such as ALIGNN, MEGNet , and CGCNN are widely accessible for predicting properties like bandgaps, formation energy, and optical absorption spectra.

To help narrow down the right tool for your project, let me know:

• Are you focusing on crystalline solids, polymers , or small molecules?
• What specific chemical/physical properties (e.g., thermal conductivity, stability, bandgap) do you need to predict?

Several major research institutions, technology companies, and specialized scientific software providers offer AI models and platforms trained on chemical properties and atomic structures for materials science . Major providers include Microsoft Research (creators of MatterGen for property-guided crystal generation), Schrödinger (offering pre-trained Machine Learning Property Prediction Panels for molecular and polymer properties), and specialized AI startups like CuspAI working alongside industry networks to model complex chemical spaces.

Key Providers and Their AI Offerings

• Microsoft Research (AI for Science): Offers MatterGen , a generative diffusion model trained across the periodic table to design novel inorganic materials based on specific chemistry, symmetry, and mechanical, electronic, or magnetic property constraints.
• Schrödinger: Provides enterprise computational platforms featuring a Machine Learning Property Prediction Panel pre-trained on extensive proprietary and literature data to forecast properties like boiling points, polymer glass transition temperatures, density, and dielectric constants.
• CuspAI: Focuses on AI-driven materials discovery, partnering with major tech and chemical groups to train deep learning architectures capable of generating novel chemical structures and evaluating real-world stability.
• Academic & Open-Source Ecosystems: Research hubs distribute specialized architectures like ALIGNN (Atomistic Line Graph Neural Network for predicting atomic/optical properties), MatSciBERT (for extracting composition-property data from literature), and polyBERT (specialized for polymer chemical language modeling) via repositories like GitHub and academic consortiums.

If you can share what class of materials (e.g., polymers, inorganic crystals, metal alloys) or specific properties (e.g., thermal, mechanical, electronic) you want to target, I can help you pick the most suitable AI framework or database.

If you are planning to use AI for materials science, there are several classes of models and providers to consider. The best choice depends on whether you want property prediction, molecular/material discovery, simulation acceleration, or a general-purpose materials foundation model.

1. Large materials foundation models (broad chemistry + physics)

These are closest to the “ChatGPT for materials” concept: pretrained on large chemical/material datasets and adapted to downstream problems.

---

2. Open scientific models widely used in academia

CHGNet

• Trained on Materials Project data and molecular dynamics trajectories.
• Designed for crystal structures and atomistic simulations.
• Useful for:
  • energies
  • forces
  • magnetic moments
  • relaxation tasks
    

ALIGNN

• Developed through the NIST JARVIS effort.
• Predicts dozens of solid-state and molecular properties.
• Uses crystal graphs with bond-angle information.
  

MACE

• A high-performance atomistic model architecture.
• Used for force fields and molecular dynamics.

M3GNet

• General-purpose materials model covering many elements.
• Often used as a baseline for materials property prediction and simulations.

---

3. Chemistry-focused foundation models

If your materials involve organic molecules, polymers, electrolytes, catalysts, or formulations, chemistry models may be more appropriate.

MIST

• Transformer models trained on molecular representations (SMILES).
• Designed for molecular property prediction.
• Covers many downstream chemical prediction tasks.

XenonPy.MDL

• A library of pretrained models for:
  • molecules
  • polymers
  • inorganic materials
  • physical/chemical properties
    

---

4. Commercial materials AI platforms

---

5. Data sources commonly used to train or fine-tune models

Most materials AI models rely on large computational chemistry databases:

• Materials Project — crystal structures, energies, electronic properties
• NIST JARVIS — materials properties, DFT calculations, ML benchmarks
• QM9 — small molecules
• Open Quantum Materials Database — calculated materials properties

---

Practical recommendations by application

Battery materials

• MatterSim
• CHGNet
• M3GNet
• Materials Project–trained models

Catalysts

• CHGNet / MACE
• molecular foundation models
• DFT surrogate models

Polymers

• XenonPy
• Schrödinger MS Informatics
• chemistry transformers

New inorganic compounds

• MatterSim
• ALIGNN
• JARVIS-based models

Formulations and mixtures

• Schrödinger MS Informatics
• chemistry foundation models such as MIST

---

A common industrial workflow today is:

DFT/experimental database → pretrained materials model → fine-tuning on proprietary data → active learning loop → lab validation

For a company or research group starting from scratch, I would usually evaluate MatterSim/CHGNet/MACE for atomistic materials, and MIST/XenonPy/Schrödinger tools for molecular and polymer systems.

If you tell me your target area (battery, semiconductor, polymer, catalyst, alloy, ceramics, etc.) and whether you need open source or commercial, I can narrow this to the best models.

Yes. The market is increasingly split between commercial materials-AI platforms and open/foundation models. The right choice depends heavily on whether you mean molecular properties, crystal properties, or process/formulation properties.

Strong options to look at

Schrödinger is probably the most mature commercial choice if you're looking for something resembling a materials-science modeling stack: its MS Informatics product covers organic molecules, polymers, formulations, organometallics and inorganic solids, with pretrained ML models and physics-informed descriptors.

Citrine is especially interesting if your eventual plan is to train models on proprietary experimental data. Its VirtualLab lets researchers select target properties, build models from chemical/material/process variables, and search for formulations that meet specifications. Citrine says its system is designed to work even with relatively small experimental datasets.

For an open research route, MatterGen is one of the most interesting models. It generates inorganic materials and can be conditioned on target properties; Microsoft explicitly supports fine-tuning it on your own property-labeled materials.

GNoME is somewhat different: it is primarily a crystal-stability/discovery model, rather than a general-purpose "give me any chemical property" model. Google used it to predict 2.2 million crystal structures, with about 380,000 predicted to be particularly stable, and released the resulting structures/data to the research community.

An important distinction

If by "trained on chemical properties" you mean something like:

Input: molecular structure / SMILES / crystal structure
Output: band gap, formation energy, melting point, conductivity, modulus, adsorption energy, solubility, etc.

then I'd look at three classes of models:

1. Molecular foundation models — best for small molecules, polymers and chemistry.
2. Crystal/materials foundation models — best for inorganic solids and atomistic properties.
3. ML interatomic potentials (MLIPs) — best when you want energies, forces, structures and molecular dynamics at a fraction of the cost of DFT.

The Materials Project is particularly valuable as infrastructure here: its ML datasets cover electronic, thermal, thermodynamic and mechanical properties, and it explicitly supports training models from chemical composition and/or structure.

If you're building an AI-for-materials program

I'd shortlist them this way:

• Want to use AI commercially with experimental data → Citrine
• Want an integrated computational chemistry/materials platform → Schrödinger
• Want to generate new inorganic materials → MatterGen
• Want large-scale inorganic-crystal discovery → GNoME + Materials Project
• Want AI-accelerated atomistic simulation → Orbital / MLIP ecosystem
• Want to build your own models → Materials Project + MatBench + open foundation models

One caveat: these aren't necessarily models that have learned experimental chemical reality. A large fraction of materials-AI training data comes from DFT/computational labels, and recent work emphasizes that models trained primarily on composition and idealized structures can struggle with processing, microstructure, synthesizability, and real-world performance.

If you tell me what kind of materials you're working with (e.g. polymers, batteries, catalysts, semiconductors, alloys, ceramics, coatings), I can narrow this to the 5–10 models/platforms that are actually relevant, including APIs, pricing/access, training data, and whether you can fine-tune them on your own data.

Several major tech companies, specialized scientific software providers, and research institutions offer AI models specifically pre-trained on chemical properties, crystal structures, and molecular data for materials science:

• Microsoft Research developed MatterGen , a generative AI model trained to design new inorganic materials from scratch based on specific property requirements (such as stability or targeted electrochemical behaviors). They also integrate advanced AI-driven property models and machine-learned force fields into Azure Quantum Elements.
• CuspAI works on foundational AI models for materials discovery, partnering with heavyweights like Meta and Nvidia to build scalable screening and predictive systems that map out functional chemical and structural spaces.
• Schrödinger offers commercial computational suites featuring a Machine Learning Property Prediction Panel . These pre-trained models predict key chemical and physical properties (like boiling point, vapor pressure, density, viscosity, and glass transition temperature) using curated proprietary and literature datasets.
• The Materials Project & Open-Source GNN Ecosystem provides open access to massive property datasets and pre-trained graph neural networks (GNNs) like ALIGNN, MEGNet , and CGCNN . These models are trained on quantum-mechanical properties (via Density Functional Theory calculations) to predict properties like formation energy, band gaps, and elastic tensors.

To help narrow down the right model or platform, what specific class of materials (e.g., polymers, crystalline inorganic solids, metal-organic frameworks) and what properties (thermal, electronic, mechanical) are you primarily trying to predict or design?

Yes. There are now several AI models and platforms that are specifically pre-trained on chemical structures, molecular properties, and materials datasets rather than only general text. The best choice depends on whether your work is focused on molecules, crystalline materials, polymers, catalysts, or atomistic simulations.

Here's a comparison of the main options:

IBM Research

IBM has released several chemistry-specific foundation models, including MoLFormer and SMI-TED. They are trained on tens of millions of molecules represented as SMILES strings and support:

• molecular property prediction
• similarity search
• molecular embeddings
• candidate generation
• fine-tuning for downstream chemistry tasks

IBM's materials models also include multiple representations (SMILES, SELFIES, molecular graphs, and 3D structures), making them useful beyond drug discovery.

MIST (University of Michigan + Argonne)

MIST (Molecular Insight SMILES Transformers) is one of the strongest open-source options.

Highlights:

• pretrained on extremely large molecular datasets
• fine-tuned for roughly 400 chemical property prediction tasks
• predicts
  • thermodynamic properties
  • electrochemical properties
  • quantum chemistry properties
  • mixture properties
• designed specifically for chemistry and materials discovery

If you want a model you can fine-tune yourself, MIST is among the best current open models.

ChemFM

ChemFM is a 3-billion-parameter foundation model trained on approximately 178 million molecules.

Reported strengths include:

• molecular property prediction
• reaction prediction
• conditional molecule generation
• transfer learning with relatively small labeled datasets

Its creators report strong performance across dozens of chemistry benchmarks.

Microsoft Research

Microsoft has developed foundation models aimed more directly at materials science than molecular chemistry.

These include work such as:

• MatterGen
• MatterSim

These models focus on

• crystal structures
• inorganic compounds
• batteries
• semiconductors
• catalysts

rather than pharmaceutical molecules.

Orbital Materials

Orbital Materials builds proprietary foundation models specifically for advanced materials.

Their work targets:

• carbon capture materials
• catalysts
• battery materials
• industrial chemicals

Unlike many academic models, these are commercial systems intended for industrial R&D.

CuspAI

CuspAI is building AI systems for end-to-end materials discovery.

Their platform combines:

• scientific literature
• chemistry databases
• simulation
• generative molecular design

to propose novel materials for industrial applications.

Which is best?

• Open-source chemistry research: MIST
• General molecular property prediction: IBM MoLFormer or ChemFM
• Crystal and inorganic materials: Microsoft MatterGen/MatterSim
• Industrial materials discovery: Orbital Materials or CuspAI

If you're doing academic materials science—such as batteries, alloys, polymers, catalysts, or semiconductors—a practical workflow is often to combine:

1. a general LLM (for literature review, planning, and code),
2. a chemistry foundation model like MIST or ChemFM (for molecular or property prediction), and
3. domain-specific simulation tools (such as density functional theory or molecular dynamics) to validate promising candidates. This combination leverages both the broad reasoning of language models and the specialized knowledge encoded in chemistry-focused models.

The landscape for AI models trained on chemical properties and materials science spans major tech labs, national institutions, and open-source scientific communities. These organizations offer pre-trained foundation models, specialized graph neural networks, and generative frameworks designed to predict properties, screen stability, or design new compounds.

• Google DeepMind
  • Description: Famous for releasing models like GNoME (Graph Networks for Materials Exploration), which vastly expanded the catalog of stable crystal structures and predicted the thermodynamic stability and chemical properties of hundreds of thousands of new inorganic compounds.
• Microsoft Research (AI for Science)
  • Description: Developed MatterGen , a generative AI model for materials design. It allows researchers to condition the generation of new inorganic crystals on targeted chemical, mechanical, electronic, or magnetic properties.
• IBM Research
  • Description: Active in building multi-modal foundation models for chemistry and materials science. Their initiatives focus on pre-training models across diverse chemical representations (such as SMILES text strings, molecular graphs, and 3D atomic structures) to predict properties and optimize industrial materials.
• U.S. Department of Energy (DOE) National Labs (Argonne, Lawrence Berkeley, Los Alamos)
  • Description: Institutions like Lawrence Berkeley Lab (creators/stewards of the Materials Project ) and Argonne National Laboratory develop advanced tools and open suites like MIST (Molecular foundation models for chemistry and materials discovery) and solid-state reaction pathways models.
• Academic & Open-Source Architectures (ALIGNN, CGCNN, MEGNet)
  • Description: If you prefer running open-source architectures yourself rather than using a proprietary platform, frameworks like ALIGNN (Atomistic Line Graph Neural Network) and SchNet are widely used. They are pre-trained on crystal symmetry, atomic coordinates, and quantum interactions to predict properties like optical spectra, formation energies, and alloy behavior.

If you'd like to narrow this down, tell me:

• Are you focusing on crystalline/inorganic materials, polymers , or small molecules?
• What specific chemical properties (e.g., formation energy, conductivity, spectra) do you need to predict?

I can point you toward the exact repository or toolkit best suited for your workflow.

If you're working in materials science, there are several categories of AI models worth considering. Some are general-purpose LLMs with scientific capabilities, while others are foundation models specifically pretrained on molecules, materials, or chemical properties.

Some notable chemistry- and materials-specific models include:

• MIST (Molecular Insight SMILES Transformers): An open-source family of models pretrained on very large molecular datasets. Fine-tuned versions cover hundreds of downstream property prediction tasks including thermodynamic, electronic, biochemical, and mixture properties, making them attractive for materials research.
• IBM SMI-TED: A chemical language model trained on roughly 91 million SMILES strings, intended for molecular embeddings, property prediction, and transfer learning in chemistry and materials science.
• ChemFM: A 3B-parameter chemistry foundation model pretrained on approximately 178 million molecules. It has demonstrated strong results on molecular property prediction, reaction prediction, and molecular generation benchmarks.

For atomistic simulations, another rapidly growing class of models consists of foundation machine-learned interatomic potentials (MLIPs). These are designed for molecular dynamics and materials simulations rather than text generation, and recent work focuses on models that generalize across many chemical systems while remaining computationally efficient.

Which is best depends on your goal

• Predict material properties (band gap, formation energy, elastic constants): MIST, IBM's materials models, and atomistic foundation models are among the strongest options.
• Design new molecules or materials: ChemFM, MIST, and startups like CuspAI focus on inverse design and generative discovery.
• Build your own models: Open-source models such as MIST and IBM's chemistry models are easier to fine-tune on proprietary datasets than closed commercial systems.
• Literature review and research assistance: General LLMs (such as ChatGPT or Gemini) can help interpret papers and write code, but they are not themselves pretrained specifically on chemical properties.

The field is evolving quickly. Recent surveys describe a shift toward multimodal foundation models that combine molecular graphs, crystal structures, spectroscopy, simulations, and text into unified models for materials discovery.

If you can share your specific area—such as batteries, catalysts, polymers, semiconductors, alloys, or pharmaceuticals—I can recommend the models and datasets that are most widely used for that subfield.

Using AI for materials science and property prediction involves a mix of tech giants, specialized startups, and open-source scientific initiatives. These entities train machine learning models, graph neural networks (GNNs), and diffusion models on massive quantum simulation and crystal-structure databases (like the Materials Project) to predict chemical, electronic, and mechanical properties.

The key providers and platforms offering models trained on chemical and material properties include:

• Microsoft Research (AI for Science) : Developed MatterGen , a generative diffusion model trained to design novel materials tailored to specific chemical, mechanical, electronic, and magnetic constraints. They also back MatterSim for atomistic simulations [1.22].
• Google DeepMind : Famous for launching GNoME (Graph Networks for Explorers), an AI tool that predicted the stability of over 2.2 million new crystal structures, massively expanding humanity's known stable inorganic materials catalog.
• NVIDIA : Offers the NVIDIA AlchemI ecosystem and accelerated infrastructure for chemistry and materials science, partnering with specialized firms to run batched density functional theory (DFT) and machine-learned interatomic potentials.
• CuspAI : A prominent startup focused on AI-powered materials discovery that uses generative algorithms to virtually screen and invent new materials for carbon capture, energy storage, and industrial chemistry.
• Entalpic : An AI-first platform that designs molecules and materials for industrial and surface-driven processes using atomistic AI models and high-throughput quantum simulation feedback loops.
• Academic & Open-Source Hubs : Projects like the Materials Project provide massive open datasets and API-accessible machine-learning toolkits (such as matgl ) for property prediction. Additionally, tools like MIT's Crystalyze offer specialized models for identifying crystalline structures from experimental powder data.

If you share what specific property you are trying to predict or optimize (e.g., band gaps, crystal stability, polymers, or catalysis), I can point you toward the best open-source framework or specific model architecture for your project.