LabHub

Blog

AI for Materials Science and Chemistry 2026 Deep Dive — DeepMind GNoME, Microsoft MatterGen, Orbital Materials, Citrine Informatics, Schrödinger, RoseTTAFold All-Atom, Boltz-1, MACE, NequIP, DPMD

2026-05-16 · 20 min read · #ai-materials#chemistry-ai#gnome-deepmind#mattergen-microsoft#orbital-materials#citrine-informatics#schrodinger#rosettafold-aa#boltz#mace#nequip#dpmd#english

한국어English日本語

Opening — May 2026, materials science is passing its AlphaFold moment

The 2024 Nobel Prize in Chemistry going jointly to David Baker (de novo protein design) and Demis Hassabis with John Jumper (AlphaFold) was not just a news item. It was a declaration that computational chemistry plus deep learning is now an academically validated research methodology. The same wave is spreading beyond proteins to inorganic crystals, catalysts, battery electrodes, and polymers.

Once DeepMind GNoME announced 2.2 million stable crystal candidates in November 2023, Microsoft published the generative crystal model MatterGen in January 2025, and Orbital Materials open-sourced the GPT-style materials foundation model Orb-v3 in the same quarter, the table flipped. As of May 2026, the estimate that the average cost of discovering one new material has dropped to one ten-thousandth of a decade ago is being seriously debated. Following the iter86 drug-design post, this article zooms into the adjacent territory of materials science and chemistry.

Why materials AI matters in 2026 — the collapse of discovery cost

Traditional materials science ran on the guess, synthesize, measure loop and averaged 18 to 30 years per generation. Lithium-ion batteries, perovskite solar cells, and high-temperature superconductors all went through similarly long discovery-to-commercialization cycles. The AI attempt to break this loop combines three tools.

By chaining these three, the inverse-design pipeline has become standard: generate candidates from a property target, evaluate quickly with MLIPs, verify with DFT, then synthesize.

DeepMind GNoME — the shock of 2.2M crystal candidates

GNoME (Graph Networks for Materials Exploration) is the crystal-discovery project DeepMind published in Nature in November 2023. The core result fits in one line.

The shock was not just a number. The known stable inorganic crystals grew from about 48,000 (everything humanity had catalogued by 2023) to 2.2 million — two orders of magnitude in a single step. In a follow-up, Berkeley A-Lab autonomously synthesized 41 of the GNoME candidates in 2024, demonstrating the AI-discovery plus autonomous-synthesis loop end to end.

DeepMind kept the GNoME model itself closed and released only the data. Open models built on the same idea by follow-up researchers have appeared rapidly since.

Microsoft MatterGen — the standard for generative crystal design

Released in Nature and open-sourced simultaneously by Microsoft Research in January 2025, MatterGen is a diffusion-based generative model for crystal structures. Where GNoME discovered stable crystals, MatterGen designs crystals that meet target properties.

Microsoft also released MatterSim on the same line. MatterSim acts as a fast property evaluator for structures generated by MatterGen, so the two together form a vertically integrated generate-then-evaluate pipeline from a single vendor.

The MatterGen GitHub repository is MIT-licensed.

# MatterGen inference example (pseudocode from official README)
from mattergen.evaluation import StructureValidator
from mattergen.generation import MatterGenPipeline

pipe = MatterGenPipeline.from_pretrained("microsoft/MatterGen")
structures = pipe.sample(
    composition={"Li": 1, "Fe": 1, "P": 1, "O": 4},
    target_bandgap=2.5,
    n_samples=100,
)

validator = StructureValidator.from_pretrained("microsoft/MatterSim")
scored = validator.score(structures)

Orbital Materials Orb-v3 — open-source GPT-style materials foundation model

The UK startup Orbital Materials rose quickly in 2024-2025 by releasing the Orb series. Orb-v3, shipped in late 2025, has the following traits.

Orbital Materials publicly states the goal is a "GPT-3 moment for materials science": train one large foundation model, then fine-tune for diverse downstream tasks — the same pattern applied to materials.

Similar lineups include DPA-2 (DeepModeling consortium, China), EquiformerV2 (MIT, on Hugging Face), and PreFerredPotential by PFN inside Matlantis. As of May 2026 there are five to six competing materials foundation models.

MACE — accuracy leader of machine-learned interatomic potentials

The most expensive part of molecular dynamics (MD) is computing interatomic forces at every time step. Direct DFT puts a single simulation into many days. Machine-Learned Interatomic Potentials (MLIPs) replace this force calculation with a neural network and accelerate four to six orders of magnitude.

MACE (MACE for Atomistic Simulations), built by the ACEnet group at Cambridge, is an equivariant MLIP regarded as the accuracy leader as of May 2026.

MACEs performance is verified on SPICE, MPtrj (Materials Project training split), OC22 and OC20 (Open Catalyst), and Matbench, where it consistently lands in the top ranks. The foundation-model variants (MACE-MP-0 and MACE-MP-2) ship pretrained weights you can download and apply zero-shot to diverse systems.

# MACE Foundation Model usage example
from mace.calculators import mace_mp
from ase.build import bulk

atoms = bulk("Si", "diamond", a=5.43)
calc = mace_mp(model="medium", dispersion=False, default_dtype="float32")
atoms.calc = calc
energy = atoms.get_potential_energy()
forces = atoms.get_forces()

NequIP and Allegro — the other axis of equivariant neural potentials

Alongside MACE, the other major equivariant-MLIP lineage is NequIP (Neural Equivariant Interatomic Potentials) and its successor Allegro. Both come from the Harvard-MIT joint group of Boris Kozinsky.

Given the same training data, MACE and NequIP plus Allegro often tie on accuracy. The difference shows up in hardware affinity and training stability. On very large systems with tens of thousands of atoms, Allegros shallow message-passing depth tends to be faster.

DeePMD-kit — the large-scale MLIP framework with Chinese roots

DeePMD-kit (DPMD) is the MLIP framework developed in collaboration between the Beijing Institute of Computer Science and Princeton e-CAS. The follow-up maintenance is handled by the DeepModeling consortium (Bohrium, AISI Beijing, and others).

Comparable to LLaMA on the LLM side, DPMD is the "eastern axis" of materials-MLIP. Citations and GitHub stars stand shoulder to shoulder with MACE and NequIP.

SchNet, PaiNN, MatGL, CHGNet — the first-generation graph neural potentials

Before equivariant MLIPs arrived, the standards for materials GNNs were SchNet and PaiNN, both from the TU Berlin group. In 2026 they are still used for lightweight systems and fast prototypes.

CHGNet is often cited as a candidate "universal crystal simulator" because it covers every magnetic and non-magnetic structure in the Materials Project training set.

Materials Project — emperor of the public databases

Materials Project (Lawrence Berkeley National Lab plus UC Berkeley, Kristin Persson) is a public DFT database for inorganic crystals launched in 2011. As of May 2026 it publishes DFT calculation results for roughly 150,000 inorganic crystals.

# Materials Project API usage example
from mp_api.client import MPRester

with MPRester(api_key="YOUR_KEY") as mpr:
    results = mpr.materials.summary.search(
        elements=["Li", "Fe", "P", "O"],
        formula="LiFePO4",
        fields=["material_id", "formula_pretty", "band_gap", "formation_energy_per_atom"],
    )
    for r in results:
        print(r.material_id, r.formula_pretty, r.band_gap)

The 2.2 million candidates GNoME discovered have been progressively integrated into MP since 2024. In other words, MP is also the place to query GNoME results.

OQMD, AFLOW, NOMAD, JARVIS — four more big public databases

Beyond Materials Project there are four more public inorganic-materials databases.

Each database has a different strength. For alloy searches use OQMD, for 2D materials use JARVIS, for multi-code compatibility use NOMAD, and for general entry the usual choice is Materials Project.

Citrine Informatics — the enterprise materials-AI standard

Citrine Informatics (founded 2013, headquartered in California) sells materials-science AI to the enterprise. As of May 2026 it is the R&D partner of dozens of global chemicals and materials companies including BASF, Panasonic, Boeing, and Saint-Gobain.

Citrine focuses less on "AI-driven discovery" and more on decision automation: gather all internal data a company already owns (patents, lab notebooks, simulation results) and turn it into decision models.

Competitors with similar positioning include Materials Zone (Israel), Uncountable (United States), and Phaseshift (United States).

Schrödinger — the NASDAQ-listed pharma and materials heavyweight

Schrödinger (NASDAQ: SDGR) started in 1990 as a quantum-chemistry software company. As of May 2026 it is a listed company with a market cap around 8 billion dollars and runs drug discovery plus materials discovery as two parallel tracks.

Schrödinger advances its own drug candidates (SGR-1505, SGR-2921, and others) into clinical trials while signing collaborative R&D contracts with big pharma such as BMS, J&J, and Takeda. A large share of revenue comes from software licensing plus drug-collaboration milestones.

RoseTTAFold All-Atom — unifying proteins and materials in one model

Published in Science in May 2024 by David Bakers group at the University of Washingtons Institute for Protein Design, RoseTTAFold All-Atom (RFAA) predicts proteins, nucleic acids, small molecules, metal ions, and covalent modifications in a single model. It is the successor to RoseTTAFold 2, which only handled proteins.

David Baker shared the 2024 Nobel Prize in Chemistry for the RoseTTAFold plus RFdiffusion plus protein-design paradigm. The iter86 drug-design post covered the protein side in more depth.

Boltz-1, Chai-1, ESM-3 — the open alternatives to AlphaFold 3

Right after AlphaFold 3 appeared in Nature in May 2024, the open-source alternatives showed up fast.

As of May 2026, Boltz-2 is the most-cited AF3 alternative in academia. Chai-1 is more widely adopted in commercial drug-discovery workflows.

Atomwise, Isomorphic Labs, Genesis Therapeutics — two flavors of AI drug discovery

The AI drug-discovery companies that sit on the border of materials science and molecular design deserve a separate roundup. As a complement to the iter86 drug-design post, here is the short list.

These are AI-drug-discovery shops at heart, not materials companies, but they share the molecular-simulation plus graph-neural-network plus generative-model stack.

DFT plus AI — VASP, Quantum ESPRESSO, JAX-DFT, DM21

Classical DFT remains the backbone of materials science. AI does not replace it; it accelerates it.

ASE (Atomic Simulation Environment, Python) wraps all of the above behind a common interface. Doing materials science in 2026 without ASE is hard to imagine.

Synthesis-route planning — IBM RXN, AiZynthFinder, Synthia, Postera

Once you find a candidate drug or material, the next step is the synthesis route: starting materials, reactions, and pathway predicted by AI.

Retrosynthesis-model accuracy is considered close to the level of expert chemists as of 2026. AiZynthFinder is open source, so it is the most common citation in academic work.

# AiZynthFinder usage example
from aizynthfinder.aizynthfinder import AiZynthFinder

filename = "config.yml"
finder = AiZynthFinder(configfile=filename)
finder.stock.select("zinc")
finder.expansion_policy.select("uspto")
finder.target_smiles = "Cn1cnc2c1c(=O)n(C)c(=O)n2C"
finder.tree_search()
finder.build_routes()
print(finder.routes[0].metadata)

Automated synthesis labs — Strateos, Emerald Cloud Labs, OpenTrons

Once AI recommends candidates and lays out synthesis routes, what remains is actually making the compounds. In 2026, automated synthesis labs fall in two camps.

A-Lab autonomously synthesized 41 of the GNoME candidates in 17 days (Nature 2023). This loop — AI discovers, robots synthesize, data feeds back — is the missing piece to close a self-improving research cycle.

Battery and energy materials — solid state plus molecular simulation

Batteries are the largest application area for materials AI. The most-watched threads:

Batteries demand "materials plus cell design plus cycle life plus safety" all at once, so MLIPs alone are not enough. The integrated electrochemistry plus thermodynamics plus mechanics simulation is what Citrine, Materials Zone, and friends compete on.

US Materials Genome Initiative and EU Battery 2030+

Government-scale materials-AI programs are most clearly laid out in the United States and the EU.

The shared core is data standards plus open tooling plus academy-industry partnership. The US tends to ship the data and tools, while the EU keeps the same toolkit but leans harder into industrial transformation.

Korean materials AI — KIST, KIMS, POSTECH, Samsung, LG

The Korean materials-AI ecosystem stacks along these lines.

The Korean government has materially increased the budget for the Korean Materials Genome Initiative since 2023. As of May 2026, the KMGI data platform (MaPS) is in beta operation.

Japanese materials AI — NIMS, PFN Matlantis, AIST, Mitsubishi Chemical

Japan stands out for the tight coupling of government, big corporates, and academia.

Matlantis deserves attention. It is the first global service to offer MLIPs as a commercial SaaS rather than as an academic toolkit. Chemicals and materials companies file simulation requests via API and receive results back.

Cheminformatics libraries — RDKit, DeepChem, PyG, e3nn

The four standard Python libraries for molecules and crystals are these.

# RDKit fingerprint example
from rdkit import Chem
from rdkit.Chem import AllChem

mol = Chem.MolFromSmiles("CC(=O)OC1=CC=CC=C1C(=O)O")  # aspirin
fp = AllChem.GetMorganFingerprintAsBitVect(mol, radius=2, nBits=2048)
print(f"fingerprint bits set: {sum(fp)}")

Conferences and workshops — MRS, ACS, NeurIPS AI4Mat

The main venues where materials-AI research lands:

The trend is clear. Materials tracks at pure-ML conferences (NeurIPS, ICML) are growing fast while traditional materials conferences (MRS, ACS) are folding ML sessions into the main program.

Limits and open challenges — what AI materials science still cannot solve

As of May 2026 the following limits still apply.

The pace at which these limits dissolve determines how far the "materials AlphaFold moment" can go.

Closing — May 2026, the era of guessing is ending

The motivating question was simple: "Can materials science have its AlphaFold moment?" The May 2026 answer is "it is already happening". GNoMEs 2.2 million candidates, MatterGens conditional generation, Orb-v3s open foundation, MACE and NequIP MLIPs, the Materials Project infrastructure, and the A-Lab autonomous synthesis robot — all of these pieces came together inside two or three years.

The remaining job is connecting these tools to industrial transformation. Real battery cells, real cancer drugs, real OLED panels — not just papers. That makes commercial bridges like Citrine, Schrödinger, and Matlantis the pivotal variable for the next five years.

If you work in materials science, this is the most volatile moment in years. It is also the highest-reward moment. The people who win fastest are those who layer AI tools on top of their domain knowledge instead of recoiling from them.

References

Comments

No comments yet.

Sign in to leave a comment