- Published on
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
- Authors

- Name
- Youngju Kim
- @fjvbn20031
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.
- DFT (density functional theory) simulation: first-principles quantum mechanics. Accurate but takes hours per calculation.
- Machine-learned interatomic potentials (MLIP): keep DFT accuracy while accelerating inference 10^4 to 10^6 times.
- Generative models: create new crystal structures conditioned on target properties.
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.
- 2.2 million new stable crystal candidates, of which 380 thousand are classified as highly likely to be synthesizable.
- Active-learning loop: a graph neural network proposes candidates, DFT verifies them, verified results are fed back into training.
- Released: the data was integrated into the Materials Project Database.
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.
- Conditional generation: chemistry composition, symmetry, magnetism, bandgap, and elastic moduli are all input conditions.
- Diffusion architecture: noise added to crystal coordinates plus lattice parameters, then reverse-diffused.
- Equivariant GNN backbone: rotation and translation symmetry are baked into training.
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.
- GPT-style decoder architecture: autoregressive generation of atomic sequences.
- Trained on hundreds of millions of DFT simulation snapshots.
- Handles property prediction, structure generation, and molecular dynamics simulation in one model.
- Apache 2.0 open source: weights are published on the Hugging Face Hub.
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.
- Higher-order many-body message passing: directly models 2-, 3-, and 4-body interactions.
- E(3) equivariance: rotational and translational symmetries are architecturally guaranteed.
- Multi-fidelity: trains well on relatively small DFT datasets.
- MIT license: published on the Cambridge ACEnet GitHub.
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.
- NequIP: equivariant message passing. Large accuracy gain over SchNet on small datasets.
- Allegro: NequIPs successor. Non-chaining local structure gives better GPU parallelism.
- Both are MIT-licensed open source.
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).
- Deep Potential Smooth Edition (DeepPot-SE): smooth encoding of atomic coordinates.
- DPA-1 and DPA-2 foundation models: pretrained on hundreds of thousands of molecular and crystal systems.
- LAMMPS integration: plugs into the existing LAMMPS MD simulator.
- DP-GEN: automation tool for the active-learning loop.
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.
- SchNet: distance-based message passing. Proposed in 2017, applied to both molecules and crystals.
- PaiNN: directional extension of SchNet. Faster and more accurate.
- MatGL and M3GNet: UC San Diego, officially backed by Materials Project. Unified model for crystals plus molecules.
- CHGNet: UC Berkeley LBNL. Learns magnetic states and charge distributions as well.
- ALIGNN: NIST JARVIS lineup. Uses line-graph encoding.
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.
- Properties served: formation energy, bandgap, elastic constants, magnetic state, dielectric constant, and dozens more.
- REST API: directly queryable via the
mp-apiPython library. - MPContribs: a user-contribution track.
- The training data for MatGL, M3GNet, and CHGNet mostly comes from MP.
# 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.
- OQMD (Open Quantum Materials Database): Northwestern University. More than one million DFT calculations, weighted toward alloys.
- AFLOWlib: Duke University. Automated DFT workflow results, tightly coupled with the Inorganic Crystal Structure Database (ICSD).
- NOMAD: Max Planck plus the EU FAIR-DI consortium. Stores results from every DFT code in a unified format.
- JARVIS (Joint Automated Repository for Various Integrated Simulations): NIST. Covers crystals, 2D materials, magnetism, topology, and more.
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.
- Citrination platform: data ingestion, active learning, candidate recommendation, and experiment design.
- Pythonic SDK: the
citrine-pythonpackage lets you define data models and design experiments programmatically. - Offered both on-premises and SaaS to fit the security postures of chemicals and materials companies.
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.
- Maestro: GUI environment for molecular modeling.
- Jaguar: quantum chemistry calculation engine (DFT, MP2, and more).
- Materials Science Suite: polymers, OLEDs, battery electrolytes, and more.
- LiveDesign: cloud collaboration environment for drug discovery.
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.
- A single model handles protein-drug docking, nucleic-acid-protein complexes, and antigen-antibody complexes.
- Contemporaneous with AlphaFold 3: where AF3 is restricted to non-commercial use, RFAA allows broader academic and non-commercial use.
- RoseTTAFold Diffusion (RFdiffusion): a protein-generation model that designs new binders.
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.
- Boltz-1 (MIT, November 2024): protein-ligand complex prediction comparable to AlphaFold 3. MIT-licensed.
- Boltz-2 (late 2025): adds affinity prediction.
- Chai-1 (Chai Discovery, 2024): an open-weight model from a for-profit startup.
- ESM-3 (EvolutionaryScale, spun out of Meta FAIR): a protein sequence-structure-function LLM.
- OpenFold: open re-implementation of AlphaFold 2. Columbia plus AQEMIA collaboration.
- ColabFold: the popular wrapper that runs AlphaFold and RoseTTAFold inside Google Colab.
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.
- Isomorphic Labs (spun out in 2021, Alphabet portfolio): drug-discovery company spun out of DeepMind. The AlphaFold 3 core team joined. Has collaborative deals with Novartis and Eli Lilly.
- Atomwise (founded 2012, San Francisco): docking with the AtomNet convolutional network. Dozens of clinical candidates.
- Genesis Therapeutics (Stanford spinout, 2019): the GEMS platform, graph-neural-network molecular design. Collaboration with Genentech.
- Recursion Pharmaceuticals (NASDAQ: RXRX): phenotypic screening plus AI. NVIDIA invested 50 million dollars.
- Exscientia (acquired by Recursion after NASDAQ merger): automated drug design.
- Insitro (founded 2018 by Daphne Koller): ML combined with induced pluripotent stem cells (iPSC) for new drugs.
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.
- VASP (Vienna Ab initio Simulation Package): proprietary. The standard at companies and national labs.
- Quantum ESPRESSO: GPL open source. The standard in academia.
- Gaussian, ORCA, Psi4: molecular quantum chemistry.
- CP2K, ABINIT: crystal and surface simulation.
- JAX-DFT (Google DeepMind, 2024): DFT implementation friendly to automatic differentiation.
- DM21 (DeepMind, 2021): a machine-learned exchange-correlation functional that improves accuracy.
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.
- IBM RXN for Chemistry: transformer-based retrosynthesis as IBM Cloud SaaS.
- AiZynthFinder (AstraZeneca, open-sourced 2020): Monte Carlo tree search plus neural networks. MIT-licensed.
- Synthia (formerly SciFinder Synthia): commercial retrosynthesis from CAS.
- Postera Manifold: born from the COVID-19 Moonshot project. Commercial SaaS.
- Molecule.one: Polish startup. Synthesizability scoring and route prediction.
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.
- Cloud labs: write the experiment as code and the remote facility runs it. The leaders are Strateos (formerly Transcriptic) and Emerald Cloud Lab.
- Benchtop automation: hardware that lives in your lab. OpenTrons (open-hardware liquid handling) and Chemspeed (commercial).
- Autonomous synthesis robots: AI decision-making plus robotic arms. Berkeley A-Lab and the University of Liverpools mobile robotic chemist are the headliners.
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:
- Solid-state electrolytes: A123 Systems, QuantumScape, Solid Power. AI is unearthing new solid-electrolyte candidates.
- Form Energy: iron-air long-duration storage. ML is used aggressively for material-candidate screening (iter96 energy post).
- DeepMind battery collaboration: 2024 LBNL collaboration on crystals and electrolytes using GNoME results.
- Toyota Research Institute (TRI): ML-accelerated battery discovery program since 2017 in partnership with UC Berkeley, Stanford, and MIT.
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.
- US MGI 2.0 (announced 2021): led by NIST. Data standardization, AI/ML integration, and workforce development.
- NIST JARVIS and MGI Hub: the data and tool infrastructure under MGI.
- EU Battery 2030+: announced 2020 as a roadmap to accelerate battery R&D by 2030.
- EU NOMAD Lab: pan-EU DFT data unification.
- EU AI4Industry: industrial-AI consortium.
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.
- KIST AI Research Institute: AI materials discovery track, with battery and catalyst applications.
- Korea Institute of Materials Science (KIMS): alloys, magnetics, ceramics. Leads the Korean Materials Genome Initiative (KMGI).
- POSTECH Department of Materials Science: materials-AI track, including the Han Yong Kim group.
- KAIST Department of Chemical and Biomolecular Engineering plus Department of Materials Science: molecule and battery ML research.
- Samsung SDI, LG Energy Solution, SK On: battery-materials AI. Their own data plus partnerships with Citrine and Schrödinger.
- Samsung Advanced Institute of Technology (SAIT): semiconductor and display new-materials AI.
- LG Chem Future Technology Center: polymer and bio-chemistry AI.
- SK hynix HBM materials-AI track: optimization of HBM interconnects and thermal-interface materials.
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.
- NIMS (National Institute for Materials Science): Japans materials-research center, established in 1956. Leads the MI2I (Materials Integration by Network Initiative) program.
- AIST (National Institute of Advanced Industrial Science and Technology): the AI4Materials consortium.
- University of Tokyo, Kyoto University, Tohoku University: materials informatics is now an established academic field.
- Mitsubishi Chemical Holdings plus Preferred Networks (PFN): announced Matlantis as a joint venture in 2021. PFNs PreFerredPotential foundation model is offered as a cloud SaaS.
- Asahi Kasei, Showa Denko, Toray, Sumitomo Chemical: each runs in-house ML materials groups.
- Toyota Research Institute (TRI, in California): a Japanese-capital US-based AI-materials outpost.
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: BSD-licensed. Parses molecule SMILES and SDF, generates fingerprints, and produces visualizations. The canonical cheminformatics toolkit.
- DeepChem: spun out of Vijay Pandes group at Stanford. Integrated chemistry ML workflows.
- PyG (PyTorch Geometric): general graph-neural-network library, frequently used for molecules and crystals.
- e3nn: equivariant neural-network library. The backend for MACE, NequIP, and Allegro.
- ASE (Atomic Simulation Environment): the integration standard for the simulation codes covered above.
- Pymatgen: official Python library of Materials Project. Manipulates crystal structures.
- PySCF: Python quantum chemistry. Molecular DFT and HF calculations.
# 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:
- MRS Spring/Fall Meeting: Materials Research Society, covering materials science broadly.
- ACS National Meeting: American Chemical Society, chemistry plus materials.
- APS March Meeting: American Physical Society, condensed-matter physics emphasis.
- NeurIPS AI4Mat workshop: held annually since 2022. The ML view of materials science.
- ICML AI4Science: the same field at ICML.
- AI for Science Workshop (ICLR, NeurIPS).
- MGI Annual Meeting and the Korean Materials Society spring and fall meetings.
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.
- Gap between synthesizability and stability: of the 2.2 million GNoME candidates, only dozens to a few hundred have actually been synthesized. "Stable on paper" is not the same as "actually makeable".
- Data bias: most DFT data is for simple crystals. Amorphous materials (glasses, polymers, alloys), defects, surfaces, and interfaces are underrepresented.
- Organic-material accuracy: ML accuracy for OLEDs, polymers, and drug molecules trails inorganic crystals.
- Magnetic and strongly correlated systems: DFTs own limits drag down reliability for magnetism, high-temperature superconductivity, and strongly correlated systems.
- Black-box concerns: generative models offer weak explanations for why a particular structure was recommended.
- Reproducibility crisis: full reproducibility requires releasing data, code, and hyperparameters. Open-sourcing is progressing but is not 100 percent yet.
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
- DeepMind GNoME paper: https://www.nature.com/articles/s41586-023-06735-9
- Microsoft MatterGen GitHub: https://github.com/microsoft/mattergen
- Microsoft MatterSim announcement: https://www.microsoft.com/en-us/research/blog/mattersim-a-deep-learning-atomistic-model-across-elements-temperatures-and-pressures/
- Orbital Materials Orb series GitHub: https://github.com/orbital-materials/orb-models
- MACE GitHub (Cambridge ACEnet): https://github.com/ACEsuit/mace
- NequIP GitHub (Harvard MIT): https://github.com/mir-group/nequip
- Allegro GitHub: https://github.com/mir-group/allegro
- DeePMD-kit official site: https://docs.deepmodeling.com/projects/deepmd/en/master/
- SchNetPack GitHub: https://github.com/atomistic-machine-learning/schnetpack
- M3GNet GitHub: https://github.com/materialsvirtuallab/m3gnet
- CHGNet GitHub: https://github.com/CederGroupHub/chgnet
- Materials Project official site: https://materialsproject.org/
- OQMD official site: https://oqmd.org/
- AFLOWlib official site: https://aflowlib.org/
- NOMAD Repository: https://nomad-lab.eu/
- NIST JARVIS: https://jarvis.nist.gov/
- Citrine Informatics official site: https://citrine.io/
- Schrödinger official site: https://www.schrodinger.com/
- RoseTTAFold All-Atom paper: https://www.science.org/doi/10.1126/science.adl2528
- Boltz-1 GitHub (MIT): https://github.com/jwohlwend/boltz
- Chai-1 announcement: https://www.chaidiscovery.com/blog/introducing-chai-1
- ESM-3 announcement (EvolutionaryScale): https://www.evolutionaryscale.ai/blog/esm3-release
- IBM RXN for Chemistry: https://rxn.app.accelerate.science/
- AiZynthFinder GitHub: https://github.com/MolecularAI/aizynthfinder
- Berkeley A-Lab paper: https://www.nature.com/articles/s41586-023-06734-w
- Materials Genome Initiative: https://www.mgi.gov/
- NIMS MI2I: https://www.nims.go.jp/MII-I/en/
- Preferred Networks Matlantis: https://matlantis.com/
- RDKit documentation: https://www.rdkit.org/docs/
- Pymatgen documentation: https://pymatgen.org/
- e3nn documentation: https://docs.e3nn.org/