The machine-learning interatomic potential (MLIP) field has consolidated around a recognizable set of ~50 academic groups, ~10 corporate labs, and a handful of well-funded startups. The clear shift in 2024–2026 has been toward universal/foundation potentials (UMA, MatterSim, MACE-MP, PET-MAD, Orb, SevenNet-Omni, PFP, DPA-3) and away from one-system-at-a-time GAP/SchNet-style fitting. Two methodological camps now dominate: pure-NN universal models (Csányi/Kozinsky/Smidt/Ceriotti/Meta/MSR), and physics-baseline + NN-correction stacks (Grimme/Isayev/Meuwly/DiStasio/Müller–Tkatchenko/Behler). The latter is fiercely contested in Europe and increasingly absorbed into industrial foundation-model efforts (AIMNet2/AIQM3, MPNICE, OrbNet), while the Asia-Pacific academic landscape — Hong Kong and Singapore in particular — remains comparatively under-staffed in dedicated MLIP-architecture work.

The remainder of this post profiles ~60 groups grouped by region, then synthesises field-level patterns.

Europe

Stefan Grimme — University of Bonn (Mulliken Center for Theoretical Chemistry)

The de facto standard-setter for cheap, robust quantum chemistry: DFT-D3/D4, GFN1/2-xTB, GFN-FF, CREST, r²SCAN-3c, and the new g-xTB (preprint mid-2025, analytic gradient 2026), aimed at ωB97M-V/large-basis accuracy at tight-binding cost. ~20 researchers; Leibniz-Prize and DFG-funded; van der Waals Prize 2025. Recent papers include the g-xTB preprint (Froitzheim, Müller, Hansen, Grimme 2025), the PTB AO-basis paper (JCP 2023), the D4 reference paper (Caldeweyher et al., JCP 2019), and the GFN2-xTB and CREST methodology papers. Key alumni: Bannwarth (RWTH Aachen prof), Bursch (FACCTS), Spicher (BASF), Goerigk (Melbourne prof), Pracht (Cambridge), Caldeweyher and Ehlert (the latter now at MSR AI4Science). Roadmap centres on g-xTB rollout, ML-fitted correction potentials, and ML-Δ stacks built on top of g-xTB. Software at github.com/grimme-lab (xtb, dftd4, crest, tblite, g-xtb, mctc-lib). Methodologically the most aligned European group for physics-baseline-plus-NN-correction work, and the natural baseline against which any new physics-augmented framework is benchmarked.

Gábor Csányi — University of Cambridge (Engineering / Lennard-Jones Centre)

FRS 2024; the European powerhouse of ML force fields. Originated SOAP, GAP, ACE, and MACE, and now drives the MACE-MP-0 / MACE-OFF / MACE-OMAT / MACE-MPA foundation-model lineage. Group of 15–25 researchers across Cambridge plus an MPI Mainz directorship; ERC, EPSRC, Microsoft and NVIDIA funding. Representative papers: Batatia et al., MACE (NeurIPS 2022); MACE-MP-0 (JCP 2025); MACE-OFF (JACS 2025); the design-space paper (Nat. Mach. Intell. 2024); the GPR review (Chem. Rev. 2021). Alumni who define the field: Deringer (Oxford), Bartók-Pártay and Kermode (Warwick), Ortner (UBC), Cheng (Berkeley), Schran (Cambridge faculty), Kovács (Microsoft), Veit (Darmstadt), De (BASF). Roadmap: full-periodic-table foundation models with multi-head DFT/CCSD(T) heads, equivariant transformers, uncertainty quantification, biomolecular MD via MACE-OFF, and BoostMD enhanced sampling. MACE is the single most-used MLIP framework worldwide and the obvious benchmark for any new architecture.

Jörg Behler — Universität Göttingen

Co-inventor of high-dimensional NN potentials (HDNNPs); 2007 Behler–Parrinello paper foundational. Group of ~10–15 (Heisenberg professorship, ICASEC, DFG funding). Methodologically the key figure on physics-aware NN potentials: 4th-generation HDNNPs with global charge equilibration (Ko et al., Nat. Commun. 2021), spin-dependent ACSFs, mHDNNPs for magnetism. Definitive review: Behler, Chem. Rev. 121, 10037 (2021). Software RuNNer and n2p2 are foundational. Alumni: Schran (Cambridge), Singraber (Vienna/n2p2), Morawietz (Boehringer), Eckhoff, Ko (UCSD with Shyue Ping Ong), Gastegger (TU Berlin/BIFOLD). Direction 2024–26: beyond-locality 4G+ potentials, electrochemistry/heterogeneous-catalysis MLPs, scalable training, community standards. A natural complement to dispersion-aware ML through charge equilibration and long-range physics.

Klaus-Robert Müller — TU Berlin / BIFOLD

Director of BIFOLD, the German national AI institute; ~30+ researchers in the broader ML group plus the Tkatchenko-collaborative chemistry subgroup. Co-creator of the original ML-for-quantum-chemistry program (Coulomb matrix, PRL 2012) and the SchNet/SchNetPack/sGDML/SchNOrb/PaiNN/SpookyNet/So3krates/SO3LR lineage. Recent flagship: Kabylda et al., SO3LR in JACS 2025 — pretrained NN with universal pairwise force field for biomolecular MD on 200k-atom systems. Frank et al., So3krates (Nat. Commun. 2024). Alumni: Schölkopf (MPI Tübingen), Schütt (Pfizer ML Research), Chmiela (TU Berlin/BIFOLD faculty), Unke (Google DeepMind), Sauceda (UNAM). The Müller–Tkatchenko axis is the most direct intellectual neighbourhood for explicitly fusing dispersion physics with NN architectures.

Volker Deringer — University of Oxford

Csányi alumnus; Oxford Professor of Materials Chemistry from 2025; ERC StG 2021 (AMADAS). ~10–15 group; pioneers GAP applications to amorphous/disordered materials and now drives automated MLIP construction (autoplex, with BAM Berlin) and benchmarking of foundation MLIPs. Recent: Nat. Commun. (2025) on autoplex and on amorphous Si paracrystallinity; Nat. Rev. Mater. (2025) review on amorphous materials; phase-change memory in Nat. Electron. (2023). Software: autoplex, graph-pes, GAP datasets. Rapidly rising profile and strong UK-faculty pipeline; well-suited for materials/disorder over chemistry.

Anatole von Lilienfeld — Vienna / Toronto Vector / TU Berlin

Architect of the modern ML-for-chemistry agenda: co-author of the Coulomb matrix (2012) and curator of QM9 (2014), QM7-X, QMugs. Distributed group across Vienna, Toronto (Clark Chair, CIFAR AI Chair) and Berlin (BIFOLD); ERC Consolidator "QML". Specialty: alchemical perturbation density functional theory (APDFT), FCHL/SLATM/aSLATM, Δ-machine learning. Representative papers: Chem. Rev. (2021); the alchemical-harmonic Δ-ML baseline paper (JCP 2025). Alumni: Unke (DeepMind), Ramakrishnan (TIFR Hyderabad), Rupp (LIST Luxembourg), Huang (Wuhan), von Rudorff (Kassel). Direction: foundation models combining QM9-style data with alchemical interpolation; APDFT for catalysis. The lineage from which Δ-ML / baseline-corrected NN methods most directly inherit.

Michele Ceriotti — EPFL (COSMO)

~25–30 researchers; one of Europe’s largest atomistic-ML labs; NCCR MARVEL/SNSF/ERC funded. Originated SOAP (with Csányi), NICE, librascal/rascaline/featomic, PET (Point-Edge Transformer), PET-MAD universal potential and the metatensor unified tensor framework. PET-OAM-XL leads Matbench Discovery (2026). Major contributions to nuclear quantum effects, equivariant theory, descriptor completeness, and the i-PI universal force engine. Recent flagship: Mazitov et al., PET-MAD (Nat. Commun. 2025); MAD dataset (Sci. Data 2025). Exceptional alumni record: Kapil (UCL), Rossi (MPSD Hamburg), Cheng (Berkeley), Cersonsky (UW–Madison), Grisafi (CNRS/ENPC), Nigam (MIT), Ben Mahmoud (Oxford), Litman (Cambridge). Among the most prestigious atomistic-ML postdoc venues in Europe.

Markus Meuwly — University of Basel

Mid-sized SNSF-funded group (~15) focused on reactive ML potentials and ML/MM hybrids. Flagship is PhysNet (Unke & Meuwly, JCTC 2019) and the new Asparagus unified Python toolkit for PES construction (CPC 2025). Strong record on Δ-ML, transfer learning, tunneling/instanton splittings, atmospheric Criegee chemistry. Famous alumnus: Oliver Unke (DeepMind, via TU Berlin). Of the European groups, Meuwly’s "physical-chemistry priors + NN energy" framing is closest in spirit to NN-augmented semi-empirical methods, and Asparagus is a credible parallel platform to other physics-baseline frameworks.

Frank Noé — FU Berlin & Microsoft Research AI4Science Berlin

Two ERC grants; co-leads MSR AI4Science Berlin (~25–35 across FU + MSR). Pioneered Boltzmann Generators, PauliNet (NN wavefunctions), CGSchNet/CGnet coarse-graining, and now BioEmu (Lewis et al., Science 2025), a generative emulator of protein equilibrium ensembles. Direction: foundation generative models for molecular simulation; transferable BGs (Klein & Noé, NeurIPS 2024). Software: bgflow, deeptime, BioEmu (released through MSR). Adjacent to MLIPs proper, but the natural place for MD-time-scale acceleration on top of MLIPs. Alumni placement is strong (Clementi at FU Berlin, Olsson Chalmers, Schütt Pfizer, Husic Princeton).

Reinhard Maurer — Warwick → Vienna → Göttingen (Humboldt Professorship 2026)

Tully alumnus (postdoc Yale), and a long-running Tkatchenko/Müller collaborator (co-author of SchNOrb, Nat. Commun. 2019). Awarded a 2026 €5M Humboldt Professorship at Göttingen and a Vienna chair. ~15–20 across sites. Specialty: MLIPs and ML wavefunctions for surfaces and non-adiabatic dynamics, including learned electronic friction. Recent: npj Comput. Mater. 2025 review. Alumni: Westermayr (Leipzig faculty). The Humboldt move from 2026 will likely expand recruiting; among Europe’s best emerging destinations for ML-surfaces-and-dispersion work.

Bingqing Cheng — UC Berkeley (Jan 2024 onward) + ISTA visiting

Ceriotti PhD, Cambridge JRF, ISTA (2021–24), now Berkeley Chemistry/BIDMaP. CACE (Cartesian ACE; npj Comput. Mater. 2024) and Latent Ewald Summation for long-range electrostatics in MLIPs (2024–25 JCTC; arXiv 2507.14302) are her flagship recent contributions. Free energies, water phase diagrams (her Nature 2021 paper with Ceriotti is canonical), high-pressure hydrogen. Group ~8–12 split Berkeley/ISTA. Patent on LES filed by UCB. The LES framework is a universal augmentation for long-range electrostatics on top of any short-range MLIP — conceptually parallel to wrapping NN corrections around a physics baseline.

Matthias Rupp — LIST (Luxembourg Institute of Science and Technology)

Co-author of the Coulomb-matrix paper (PRL 2012). Career trajectory FHI Berlin → Konstanz → Citrine → LIST. Now leads the Process Modelling/Automation/Robotisation group at LIST (~5–10 researchers). Recent representative work in Phys. Rev. B (2025) on hydrogen liquid–liquid transition; ELLIS Fellow. Smaller and more applied than the academic centres, but geographically adjacent to Tkatchenko in Luxembourg.

Max Welling — University of Amsterdam (AMLab) & CuspAI

Geometric-deep-learning founding figure; co-author of the VAE (Kingma & Welling 2014, ICLR Test-of-Time 2024), GCN (with Kipf), EGNN (Satorras et al., ICML 2021), EDM, SEGNN, SE(3)-Transformer. AMLab is one of Europe’s largest ML labs (30+ PhDs across PIs); CuspAI raised seed $30M (2024) + Series A $100M (Sept 2025; NEA, Temasek, NVIDIA NVentures), valued ~$520M, with reported $200M extension talks at unicorn valuation. Strategic partnerships with Meta (OpenDAC), Hyundai, Kemira. Alumni: Kingma, Cohen, Kipf (DeepMind), Hoogeboom (DeepMind), Satorras (DeepMind), Köhler (CuspAI), Forré (UvA). CuspAI is the most active EU MLIP-startup hiring environment.

SISSA Trieste (Baroni / de Gironcoli) and adjacent Italy

SISSA is primarily a DFT/Quantum-ESPRESSO group (Baroni, de Gironcoli, plus Giannozzi at Udine), with a real but smaller PANNA / PANNA 2.0 Behler-Parrinello-style MLIP effort (Pellegrini, Lot, Shaidu, Küçükbenli; JCP 2023). For pure MLIP work, Mazzola (SISSA Associate Prof) and Bussi (SISSA, enhanced sampling/PLUMED) are more directly relevant; for foundation-MLIP contact, EPFL/Cambridge dominate. Küçükbenli moved Harvard → NVIDIA. SISSA is best read as a powerhouse for DFT + HPC training that increasingly integrates MLIPs rather than as a flagship MLIP destination.

North America

Boris Kozinsky — Harvard SEAS (Materials Intelligence Research) & Bosch

Joint Harvard/Bosch principal scientist; APS Fellow 2023; Gordon Bell finalist 2023. ~15+ researchers; NSF/DOE/AFOSR plus deep Bosch funding. Flagship contributions: NequIP (Batzner et al., Nat. Commun. 2022), Allegro (Musaelian et al., Nat. Commun. 2023), FLARE Bayesian active-learning MLIPs, and co-authorship of MACE-MP. Direction: scaling Allegro to biomolecules and exascale, ML for polarons/dielectric response, ionic transport in batteries. The strongest MLIP-to-industry alumni pipeline in the field: Batzner → DeepMind, Musaelian → Mirian Technologies, Lixin Sun → MSR Cambridge, Yu Xie → MSR Berlin, Owen → Lila Sciences, Falletta → Radical AI, J. Yang → Georgia Tech faculty, Kavanagh → Cambridge faculty. Software at github.com/mir-group (nequip, allegro, flare). Among the top global postdoc destinations for MLIP-trained PhDs.

Tess Smidt — MIT EECS / RLE (Atomic Architects)

Promoted to Associate Prof in 2025; AI2050 Schmidt Sciences Fellow, AFOSR YIP, DOE Early Career. ~16 group members; DOE/AFOSR/NSF/MIT-IBM/Schmidt funding. Co-developed Tensor Field Networks at Google AS, then e3nn (with Mario Geiger), the de-facto E(3)-equivariant ML framework underpinning NequIP, Allegro, MACE, DiffDock. Equiformer/EquiformerV2 (ICLR 2023/2024) is the architecture Meta adopted for OMat24 and as a UMA baseline. Recent: Symphony (3D molecule generation), Nequix (JAX rewrite of NequIP), PFT (phonon fine-tuning of MLIPs). Mario Geiger now at NVIDIA. The most central academic node in equivariant-MLIP architecture design and tightly fused with Meta FAIR Chemistry’s release pipeline.

Olexandr Isayev — Carnegie Mellon Chemistry

Carl & Amy Jones Professor; AIMNet co-architect with the late ANI-1 lineage. ~10–15 group; NSF/NIH/DOE/ONR funded. Flagships: AIMNet (2018), AIMNet2 (Chem. Sci. 2025), AIMNet2-NSE (open-shell, 2025), AIMNet-X2D, Auto3D, and most relevantly AIQM3 (Chen et al., JCTC 2026, with Pavlo Dral) — explicitly a neural-network-augmented semi-empirical hybrid targeting CCSD(T) accuracy at semi-empirical speed. AIMNet2’s design (ML short-range + physics-based electrostatics + DFT-D3) is architecturally the closest published analogue of the broader physics-baseline-plus-NN approach. Long-term ANI co-development with Roitberg (Florida) and Justin Smith (now NVIDIA). AIQM3 is the leading published benchmark for hybrid semi-empirical+NN methods. Software at github.com/isayevlab (AIMNet2, aimnetcentral, AIMNet-X2D, Auto3D).

Adrian Roitberg — University of Florida

Frank Harris Professor; PhD Buenos Aires; APS/ACS Fellow. The home of the ANI series and TorchANI lineage: ANI-1 (2017), ANI-1x, ANI-1ccx, ANI-2x (2020), ANI-1xBB reactive (2025), and TorchANI 2.0 / TorchANI-Amber (2025). NSF GOALI 2023–2028 ($4.5M) program for at-scale heterogeneous ML model development. The ANI-to-NVIDIA pipeline is the single largest industry alumni flow in MLIP: Justin Smith (Senior Devrel Manager, ALCHEMI), Xiang Gao, Roman Zubatyuk all at NVIDIA. Strong fit for biomolecular MD with NN potentials and direct industry transition.

Robert DiStasio — Cornell Chemistry

Berkeley PhD (Head-Gordon); Princeton/FHI postdocs; co-developer of the original TS and MBD vdW methods (Tkatchenko, DiStasio, Ambrosetti). Sloan Fellow 2020; Dreyfus ML award 2022; NSF CAREER. ~6–10 group. The US chemistry-side anchor of the MBD method. Notable: Ambrosetti et al., Science 2016 on wavelike charge-density fluctuations. Recent: ML-guided molecular design and ML potentials for oxides (with Benedek, Cornell MSE). Alumni at LLNL, Berkeley, faculty positions. The Tkatchenko–DiStasio axis is the methodological backbone of NN-augmented many-body dispersion work.

Mark Tuckerman — NYU Chemistry / Courant

Foundational figure in AIMD, path-integral MD, RESPA, TAMD. NSF DMR + Simons funding; ~10–15 group with Jutta Rogal sub-group. Recent: ML+path-integral MD of reactive electrolytes (JCP 2025); LLM-assisted automated MLIP dataset curation (Lahouari et al., JCTC 2026); ML-1RDM electronic structure (Nat. Commun. 2023). Software: AMLP (Automated MLIP Pipeline), MolCryst-MLIPs database. Tightly looped with Flatiron Institute. The premier US destination for combining MLIPs with rare-event sampling and ML-MD methodology.

Risi Kondor — University of Chicago

Theory anchor of equivariant ML: Cormorant, Clebsch–Gordan networks, Lorentz-equivariant NN (HEP), N-body networks. PNAS 2025 review on equivariant NN principles. Software: GElib (with Erik Thiede). Theory-first, smaller group (~5–8), but a "factory" for equivariant-ML academic talent. Direct relevance to MLIPs is via the theoretical scaffolding underlying MACE/NequIP; Erik Thiede (Cornell faculty 2022) now sits in the same building as DiStasio.

Garnet Chan — Caltech

NAS member; Bren Professor; Simons Investigator. Very large theory group (~25–35). Author of PySCF, Block/Block2, AFQMC implementations; tensor networks; quantum embedding (DMET, DMFT-style). Recent profile-defining work has been the classical-heuristic critique of quantum-advantage claims (Nat. Commun. 2023; Faraday Discussions Spiers Memorial Lecture 2024; Jan 2026 FeMo-cofactor result). ML transition is partial: ML wavefunctions inside QMC, ML-driven coarse-graining of effective Hamiltonians. Best read as a method-development laboratory producing CCSD(T)/CCSD-quality reference data and theoretical insight, rather than an MLIP shop. Highest-prestige faculty pipeline in N. American theoretical chemistry.

Berkeley/LBNL MLIP cluster

The Bay Area is now the densest US MLIP centre outside Boston. Key PIs:

  • Kristin Persson (UCB MSE / LBNL) — Materials Project director; recent foundation-MLIP dataset paper (Kaplan et al., 2025). Best for datasets and infrastructure rather than novel architectures.
  • Aditi Krishnapriyan (UCB EECS+ChemE / LBNL) — PINN failure modes (NeurIPS 2021); Gaunt Tensor Products (ICLR 2024 spotlight); MLFF distillation; deep ties to OMat24. The "ML-methods-first" complement to Cheng’s "MLIP-physics-first" group.
  • Teresa Head-Gordon (UCB) — polarizable force fields (AMOEBA/MB-pol/MB-UCB/CMM) and ML force fields with explicit polarization; Berkeley’s closest group in spirit to dispersion-aware NN methods.
  • Bingqing Cheng (profiled above; primary MLIP-architecture PI on the West Coast).

Other LBNL/Berkeley MLIP-relevant figures: Anubhav Jain (MP, Robocrystallographer), Sinéad Griffin (Molecular Foundry, ML+DFT for quantum materials), Mark Asta (alloys+ML).

Industry: Big Tech and national labs

Microsoft Research AI for Science (Cambridge UK / Beijing / Berlin / Amsterdam)

~50–80 across hubs. Key leaders Bishop (Director), Noé (Berlin), Tian Xie (Cambridge UK, MatterGen/MatterSim), Liu (Beijing), van den Berg (Amsterdam), Kruft, Lu, Shao, Wang. Flagship releases: MatterSim universal MLIP (arXiv:2405.04967, May 2024), MatterGen generative crystal model (Nature Jan 2025), AI2BMD protein-fragment ML force field (Nature 2024), Skala ML exchange-correlation functional (June 2025 preprint, OSS Oct 2025), BioEmu (Science July 2025), Aurora atmospheric model. Open source on github.com/microsoft (mattersim, mattergen, AI2BMD, skala). Hires aggressively from Csányi, FU Berlin, Welling, MIT. Commercial deployment via Azure Quantum Elements.

Meta FAIR Chemistry (Menlo Park + remote)

~15–25 core researchers led by Larry Zitnick (Director). Spun out of the Open Catalyst Project. The most consequential MLIP releases of 2024–2025 came from this team: OMat24 (DFT dataset + eqV2 inorganic-bulk MLIP, Oct 2024), OMol25 (140M ωB97M-V/def2-TZVPD calculations, 350-atom systems; arXiv:2505.08762), OMC25 Open Molecular Crystals, and UMA – Universal Models for Atoms (arXiv:2506.23971; Mixture-of-Linear-Experts on eSEN backbone; UMA-S 6.6M active params, UMA-M 50M active / 1.4B total). The OMol25 author list reads as a who’s-who of the academic MLIP world: Csányi, Batatia, Krishnapriyan, Eastman, Rosen, Rackers, Blau, Kitchin. Open source at github.com/facebookresearch/fairchem; weights gated via FAIR Chemistry License. Meta is the central hub for the academic MLIP community in 2024–2026 and the highest-yield big-tech application target for chemistry-strong PhDs.

Google DeepMind — AI for Science (London / Mountain View)

Pushmeet Kohli (VP); John Jumper (Nobel 2024); Hassabis (CEO). Materials team led by Ekin Cubuk and Amil Merchant; Simon Batzner (ex-Kozinsky) is the most public MLIP figure. Flagship: GNoME (Nature 2023, 2.2M predicted crystals, 380K stable, plus pretrained inorganic-bulk MLIP). AlphaFold 2/3, AlphaProteo, AlphaGenome, AlphaEvolve, AI Co-Scientist, AlphaEarth, GraphCast adjacent. DeepMind’s MLIP roadmap post-2023 is opaque: less open than Meta, less prolific than Microsoft. Independent critiques on GNoME novelty exist. Best for prestige + agentic-AI research with an MLIP angle, but harder to enter directly into MLIP work than Meta or MSR.

NVIDIA — ALCHEMI + BioNeMo (Santa Clara)

Strategy is substrate, not foundation models. Launched ALCHEMI at SC24 (Nov 2024); ALCHEMI Toolkit-Ops (2025) provides GPU-accelerated batched kernels for neighbor lists, DFT-D3, long-range electrostatics. NIM microservices wrap third-party MLIPs (MACE-MP, AIMNet2, OrbMolv2, UMA, eSEN). BioNeMo Framework drives the lab-in-the-loop biology stack with 200+ pharma adopters (Lilly, Genentech, Recursion, Insilico, Novartis). Key staff: Justin Smith (Senior Devrel Manager), Roman Zubatyuk, Nikita Fedik, Emine Küçükbenli, Mario Geiger. The ANI/AIMNet engineering lineage now sits at NVIDIA, making it a natural destination for chemistry-PhDs who lean systems/engineering. Anima Anandkumar departed in ~2023 back to Caltech, so the AI4Science research centre of gravity has shifted product-ward.

Industrial startups and pharma platforms

  • Orbital Materials (London/Princeton/SF). Founded 2022 by Jonathan Godwin (ex-DeepMind). Series B $200M Sept 2025 at ~$1.2B valuation; ~$220M total raised; ~50–100 employees. Internal foundation model LINUS; open-source Orb (v1/v2/v3) and OrbMol (trained on OMol25). Apache-2.0 weights. Customers: AWS, Civo, NVIDIA Inception.
  • CuspAI (Cambridge UK / Amsterdam / Berlin). Welling co-founded; ~$130M raised; ~25–30 employees. Equivariant + diffusion models for materials; MOFGEN; partners with Meta (OpenDAC), Hyundai, Kemira.
  • Iambic Therapeutics (San Diego, ex-Entos). Tom Miller (ex-Caltech) and Fred Manby. ~$300M raised; ~60 employees; OrbNet (semi-empirical-to-DFT delta learning), NeuralPLexer 1/2/3, ProPANE, Magnet. Clinical pipeline (IAM1363 in trials 2024); $2B Takeda partnership (Feb 2026). OrbNet’s symmetry-adapted atomic-orbital features are a direct intellectual cousin of NN-augmented semi-empirical methods.
  • Schrödinger Inc. (NYC, NASDAQ:SDGR). ~891 FTEs; FY2025 revenue ~$257M. MPNICE (message-passing NN with iterative charge equilibration; 89 elements; arXiv 2505.06462) is their flagship MLFF. OPLS4/OPLS5; FEP+ Protocol Builder integrating UMA in 2025-3 release. Stable industry path; under Leif Jacobson, John Weber, Karl Leswing.
  • Genesis Therapeutics (Burlingame). ~$280–327M raised; GEMS + Pearl 3D diffusion model. Lilly, Genentech, Gilead deals. Indirect MLIP fit.
  • Valence Labs / Recursion (Mila, Montreal). Released OpenQDC — ~40 QM datasets, 1.5B geometries, 70 elements, 250+ QM methods, designed explicitly for MLIP training. The most important open-science MLIP-data ecosystem outside Meta.
  • Inductive Bio, Isomorphic Labs, Atomic AI (RNA), Hummingbird Bioscience, Anthem are all adjacent and not MLIP-core.
  • Apple: no public MLIP/materials-chemistry work identified.
  • "Atomistic AI / Atomistic Inc." does not exist as a US/EU startup; the only concrete reference is ByteDance’s internal team name "Atomistic AI" used on PhD job listings.

Preferred Networks (Tokyo) — PFP/Matlantis

Founded 2014; PFP (Preferred Potential) is arguably the longest-running production universal MLIP, an equivariant GNN trained on 59M+ DFT structures, now at v7 (Sept 2024) supporting 96 elements including all naturally occurring elements plus Tc, Pm, Np, Am, Cm, with Hubbard-U mode and DFT-D3 dispersion. Core paper: Takamoto et al., Nat. Commun. 2022. Commercialised via Matlantis SaaS (joint with ENEOS); 41+ enterprise customers as of 2022. Among the world’s three or four most consequential industrial universal MLIPs, alongside MatterSim, UMA, and Orb. Closed weights/training data; only inference exposed. Hires international research engineers; PhD-track entry possible.

Asia-Pacific academic groups

Hong Kong

HKUST has a single dedicated MLIP figure, Ding Pan (joint Physics + Chemistry, Croucher Innovation Award, NSFC Excellent Young HK 2020). His Ångström group (~10) does first-principles MD + DeePMD-style MLIPs for water, ice, supercritical CO₂/water, deep-Earth aqueous chemistry, and electrochemical interfaces. Recent: deep generative model of canonical ensemble (PRL 2025); biomolecule synthesis under upper-mantle conditions (JACS 2024); hidden CO₂ kinetics in supercritical water (PNAS 2025). HKUST also operates the IAS Center for AI for Scientific Discoveries, where Pan is Associate Director; this is HKUST’s clearest hiring vehicle for an MLIP-quantum-chemistry hire. Outside Pan, HKUST Chemistry has no flagship MLIP group. Xuhui Huang has moved to UW–Madison.

CityU is materially weaker on MLIPs but receptive. The dominant figure is Xiao Cheng Zeng, recently moved from Nebraska to head the new Department of Materials Science & Engineering. h-index ~120; FRSC, AAAS, APS Fellow; recent shift toward MLFF-based phase-diagram exploration of confined water (JCP 2024). As Department Head he has direct hiring influence. The CityU Physics page lists open positions but no resident senior MLIP developer was identified.

CUHK has two relevant junior faculty: Junyi Zhu (Physics; JPCL 2024 on parameter-free electron-counting-satisfied MLIP descriptors) and Xinglong Zhang (Chemistry, joined Oct 2024; co-author of MS25 MLIP benchmark with Tibor Szilvási, JCIM 2025).

HKU hosts GuanHua Chen / ChiYung Yam / Ziyang Hu (Chemistry), whose 2025 Nat. Commun. paper "A foundation machine learning potential with polarizable long-range interactions for materials modelling" is a major HKU MLIP statement; the group also spun off HK Quantum AI Lab and MattVerse. Yue Chen (HKU Mech Eng) does MLIP-driven thermal transport.

Singapore

NTU MSE is the most active hiring environment, explicitly listing modelling and simulation among priority areas (2025). Kedar Hippalgaonkar (NTU MSE / A*STAR IMRE) anchors a S$25M materials-acceleration-platform program; ~30 group; teaches MS4672 ML-for-materials-design covering MLIPs as a topic. The NTU School of Chemistry, Chemical and Biomedical Engineering (CCEB, merged ~2024) is in growth mode with limited resident MLIP capacity.

NUS MSE has Pieremanuele Canepa (NRF Fellowship 2020–25, S$2.78M; recently took a Houston joint appointment but retains NUS-funded projects). Group of 10–15; MLIP-accelerated screening of solid electrolytes; co-author of Mach. Learn. in Materials Science (Butler/Oviedo/Canepa 2022). Recent npj Comput. Mater. 2025 with Sai Gautam (IISc) on MOFs as Zn-ion conductors using MLIPs. Strong NUS-IISc-Singapore-IHPC pipeline.

A*STAR IHPC (Benjamin Chen et al.) co-authored the 2024 J. Phys.: Cond. Matter "Roadmap for the development of MLIPs" and the MS25 MLIP benchmark — real MLIP capacity in Singapore outside academia, with joint NTU/NUS-IHPC roles common. SUTD has no atomistic MLIP presence.

Honest assessment of HK + Singapore: there is no flagship MLIP-architecture group anywhere in the region comparable to Csányi/Behler/Kozinsky, leaving substantive room for the field to grow at HKUST/NTU/CityU/NUS over the next 24 months.

China (mainland) — DeepModeling consortium

  • Weinan E (AISI Beijing + Princeton + PKU). Director of AI for Science Institute Beijing; intellectual umbrella of DeePMD/DPA/ABACUS/Uni-Mol. Funded via Beijing AI4S programs and MOST AI+Science (2025). Roadmap: OpenLAM "conquer the periodic table".
  • Han Wang (IAPCM Beijing). Lead developer of DeePMD-kit and DP-GEN; corresponding author DPA-3 (arXiv:2506.01686, 2025/26). Gordon Bell Prize 2020 (100M-atom DeePMD on Summit). IAPCM is a CAEP defense-adjacent institute, restricted for foreign students.
  • Linfeng Zhang (DP Tech / AISI). Co-founder & Chief Scientist of DP Technology (~300–500 employees; Series C 800M RMB Dec 2025; cumulative >US$200M raised). Main publications: DPA-1/2/3, DeePKS, Uni-Mol (ICLR 2023), Uni-Mol+ (Nat. Commun. 2024), Uni-Mol3 (arXiv 2025).
  • Mohan Chen (PKU). Lead developer of ABACUS ("electronic structure analysis package for the AI era," JCP 2025). Princeton PhD with Emily Carter; Gordon Bell Prize 2020. Methodologically a strong fit for fragment-MO and electronic-structure-aware MLIP development.

China — non-DeepModeling academic

  • Jun Cheng (Xiamen U / IKKEM AI4EC). Among the most globally significant Chinese MLIP-electrochemistry PIs. PhD Belfast; Cambridge postdoc (Sprik). Director of AI4EC at IKKEM. Major recent: Zhu & Cheng, ec-MLP with hybrid dielectric response (PRL 2025); Wang et al., domain-oriented universal MLP for battery electrolytes (Nat. Commun. 2025); Zhu & Cheng, computational-electrochemistry perspective (APL Comp. Phys. 2026). NSFC 22225302 + multiple major grants.
  • Hai Xiao (Tsinghua Chemistry). UCLA-Goddard PhD; Caltech postdoc. Recent flagship: AlphaNet (Yin et al., npj Comp. Mater. 2025) — local-frame-based equivariant MLIP. Also EL-MLFFs ensemble methodology. NSFC SAC center funded.
  • Yi Qin Gao (PKU). Caltech PhD with Marcus; postdoc with Voth. Very large group (~25–35 across PKU + Shenzhen Bay Lab + Changping Lab). Flagship software SPONGE / MindSPONGE (with Yi Isaac Yang) — Chinese-developed GPU MD package on Huawei MindSpore.
  • Zhi-Pan Liu (Fudan). Stochastic Surface Walking + global NN potentials; LASP is China’s main MLP ecosystem outside DeepModeling. Recent: GG-NN universal potential spanning 83 elements (Sci. China Chem. 2025).
  • Bin Jiang (USTC). EANN/REANN; gas-surface dynamics; finite-field constant-potential MLP MD (arXiv 2506.10548, 2025). The second-most-active Chinese MLIP-electrochemistry group.
  • Tarak Karmakar (IIT Delhi). Parrinello postdoc (ETH). MLIP + enhanced sampling for solution chemistry.
  • Sai Gautam Gopalakrishnan (IISc Bangalore). MIT Ceder PhD; transfer learning for migration barriers; networked with Canepa (Singapore/Houston) and Maria Chan (Argonne).
  • Sundaram Balasubramanian (JNCASR Bangalore). India’s leading MLIP-for-chemistry group; MLIPs for electrolytes, MOFs.
  • Ankit Jain (IIT Bombay MechE). MLIP-driven thermal transport.

Other notable: Niu Huang (Tsinghua) — protein-relevant MLIPs (arXiv 2601.11628, 2026); Xuefei Xu (Tsinghua) — DPMD + enhanced sampling; Shi Liu (Westlake) — ferroelectric MLIPs; Lei Li (SUSTech) — confined water; Chungen Liu (Nanjing U) — explicit electric-potential-embedded MLIP framework (arXiv 2604.07322, 2026).

China — corporate

  • ByteDance Seed for Science. The only Chinese big-tech with named, peer-reviewed MLIP work: BAMBOO for liquid electrolytes (Nat. Mach. Intell. 2025; arXiv:2404.07181) and Hi-MLIP/HINT (arXiv 2026). Open-source at github.com/bytedance/bamboo. Active hiring of ML+chemistry PhDs in Beijing/Seattle/Bay Area; Singapore office for NTU networking.
  • Tencent AI Lab / Quantum Lab, Baidu PaddleHelix, Alibaba DAMO, Huawei Pangu/MindSpore: largely adjacent (drug/protein/quantum/weather/CADD). No widely cited Chinese big-tech MLIP equivalent to DPA/BAMBOO from these four. PaddleHelix released HelixFold/HelixDock (folding/docking). Huawei’s MindSPONGE (AI-augmented MD on MindSpore) is hardware-ecosystem-focused.
  • Shanghai AI Lab (Pujiang Lab). Cui et al., "Iterative Pretraining Framework for Interatomic Potentials" (arXiv 2507.20118, 2025) — emerging foundation MLIP effort.

Japan

  • Atsuto Seko (Kyoto) — polynomial MLPs (PolyMLP), now leading the Kyoto group after Tanaka’s March 2025 retirement. Recent: PolyMLP for liquid states 22 elements (PRB 2024); P-T phase diagrams via PolyMLPs (2025). Software pypolymlp + LAMMPS user package.
  • Atsushi Togo (NIMS Tsukuba) — author of phonopy/phono3py/spglib; on-the-fly PolyMLP training for thermal-conductivity (JCP 2024 with Seko). Globally central infrastructure-level PI.
  • Isao Tanaka (Kyoto, retired March 2025) — pioneer of nanoinformatics; legacy via Seko, Togo, Y. Kumagai (Tohoku).
  • Junichiro Shiomi (UTokyo Mech Eng) — ML+materials informatics for thermal nanostructures.
  • Shinji Tsuneyuki (UTokyo Physics) — ML potentials for dielectric/optical properties; uses Matlantis PFP for La-N-H superconductor search (PRB 2024).
  • Koji Tsuda (UTokyo / RIKEN AIP) — Bayesian optimization, COMBO/PHYSBO; Ichigaku Takigawa (RIKEN AIP) — ML for catalysis.
  • Yu Kumagai (Tohoku IMR), Terumasa Tadano (NIMS) — ALAMODE anharmonic phonons.
  • Preferred Networks (industry, profiled above).

Korea

  • Seungwu Han (SNU MSE). Author of SIMPLE-NN, SPINNER, AMP², SevenNet, SevenNet-Omni. SevenNet-Omni (arXiv 2510.11241, 2025) achieves <0.06 eV adsorption-energy errors across molecules, crystals, surfaces. Recent multi-fidelity MLIP training (JACS 2025); reEWC forgetting-aware fine-tuning (npj Comput. Mater. 2026). Samsung partnership. The closest Asia-Pacific group to a real MLIP-architecture lab.
  • Yousung Jung (SNU CBE) — moved from KAIST in 2023. Berkeley PhD (Head-Gordon); Caltech postdoc (Marcus). Large group (~20+); academic partner of Meta’s Open Catalyst Project. Recent: LLM synthesizability prediction (JACS 2024, Angew. Chem. 2025); atom-to-atom mapping (Nat. Commun. 2024). More ML-for-chemistry than core MLIP.
  • Jeong Woo Han (now SNU MSE, ex-POSTECH) — MLIPs for electrochemistry/single-atom catalysis (ACS Energy Lett. 2025).
  • Sungwoo Kang (KIST), Youngho Kang (Incheon) — ex-Han alumni, SevenNet co-developers, universal MLIP fine-tuning.

Australia

  • Giuseppe Barca (Melbourne, ex-ANU). Leads the EXESS exascale electronic-structure code (frontier-scale MP2/CCSD(T)); RI-MP2 multi-GPU; CONQUEST conformer generator; Coulomb-perturbed fragmentation. JCP 2023, JCTC 2024 series. Co-founder of QDX Technologies. Bridging exascale wave-function quantum chemistry with MLIP training — particularly CCSD(T)-trained MLIPs through fragment methods — is an opening niche in the Asia-Pacific landscape.
  • Amanda Barnard (ANU) — nanoinformatics; senior editorial roles; sits on NRF Singapore CRP expert panel.
  • Asaph Widmer-Cooper (Sydney) — ML force fields for nanocrystal-ligand interfaces; AU’s most clearly MLIP-active soft-matter group.
  • Salvy Russo (RMIT), Michelle Spencer (RMIT), Debra Bernhardt (UQ) — applications and descriptor work, peripheral to MLIP development.

Strategic synthesis

Top groups to monitor most closely

  1. Stefan Grimme (Bonn) — methodological parent of D4/g-xTB and ML-Δ corrections.
  2. Olexandr Isayev (CMU) — AIMNet2/AIQM3 leads published physics-baseline+NN hybrid methods.
  3. Klaus-Robert Müller / Tkatchenko axis (Berlin/Luxembourg) — the SchNet/sGDML/SO3LR/SchNOrb lineage and explicit dispersion-aware ML.
  4. Gábor Csányi (Cambridge) — MACE foundation models are the universal benchmark.
  5. Boris Kozinsky (Harvard/Bosch) — strongest MLIP-to-industry alumni pipeline; NequIP/Allegro define equivariant MLIPs.
  6. Jun Cheng (Xiamen / IKKEM) — among the most globally significant Chinese MLIP-electrochemistry figures.
  7. Robert DiStasio (Cornell) — US chemistry-side anchor of MBD.
  8. Meta FAIR Chemistry (Zitnick) — UMA/OMol25 are setting the agenda; the 2024–26 OMol25 author network is the social graph of MLIP.
  9. Microsoft Research AI4Science (Bishop, Noé, Xie) — MatterSim/Skala/AI2BMD; well-positioned for industry transition with publishing freedom.
  10. Seungwu Han (SNU) — SevenNet-Omni is the Asia-Pacific MLIP-architecture flagship.

Methodological clusters

  • Methodologically closest groups to physics-baseline-plus-NN approaches: Grimme (Bonn), Isayev (CMU), Meuwly (Basel), von Lilienfeld (Toronto/Vienna), Müller (TU Berlin), DiStasio (Cornell). A Bonn or CMU placement most directly extends physics-baseline work; a Toronto/Vienna placement broadens into Δ-ML/alchemical learning.
  • Foundation-model groups providing breadth: Csányi (Cambridge), Ceriotti (EPFL), Kozinsky (Harvard), Smidt (MIT), MSR AI4Science, Meta FAIR Chemistry. These establish credibility as foundation-MLIP researchers.
  • Closest published methods to physics-baseline+NN hybrid (benchmarking peers): Isayev’s AIQM3 (most direct), Schrödinger’s MPNICE (most industrialised), Iambic’s OrbNet (most clinical-stage), Cheng’s CACE+LES (most foundation-model-scale), Han’s SevenNet-Omni (most cross-domain).
  • Complementary methodology groups (low overlap, high complementarity): Behler (Göttingen, charge equilibration), Tuckerman (NYU, ML+rare events), Karmakar (IIT Delhi, ML+enhanced sampling), Togo (NIMS, phonons), Persson (LBNL, datasets).
  • An emerging niche: physics-baseline-plus-NN-correction (e.g. NN-augmented xTB, D4, MBD) trained on CCSD(T)-quality data generated through fragment-based exascale quantum chemistry, deployed as drop-in replacements in production xTB/D4 ecosystems. Methodologically defensible against pure-NN universal MLIPs (UMA/MatterSim/Orb) on retrainability and physical interpretability.
  1. Foundation models have won the framing but not yet the deployment battle. UMA, MatterSim, MACE-MP, PET-MAD, Orb, SevenNet-Omni, PFP, DPA-3 are all credible universal MLIPs; none yet dominates production use because (a) inference cost remains 10–100× classical force fields, (b) coverage of charged/open-shell/biomolecular systems is uneven, and © long-range physics is handled inconsistently.
  2. Long-range physics is the open methodological frontier. Cheng’s LES, Han’s SevenNet-Omni, Behler’s 4G HDNNPs, Schrödinger’s MPNICE, and Müller’s SO3LR all converge on the same problem: how to inject electrostatics/dispersion priors into otherwise short-ranged equivariant networks.
  3. Δ-machine learning is making a comeback. AIQM3 (Isayev/Dral), Krug et al. (von Lilienfeld 2025), MPNICE, and OrbNet all use a physics baseline plus NN correction. After ~5 years of pure-NN dominance, the pendulum is swinging back toward hybrid approaches.
  4. Industry consolidation is real. Orbital, CuspAI, Iambic, Schrödinger, PFN have collectively raised >$1B in 2024–25; Meta + MSR + NVIDIA + Google are spending order-of-magnitude more on internal teams. ANI/AIMNet → NVIDIA; NequIP/Allegro → Bosch/DeepMind/MSR; Csányi → MSR/Cambridge dual; Welling → CuspAI. The academic-to-industry pipeline is now the dominant career trajectory in MLIP.
  5. Geographic shift toward Asia is uneven. China leads on volume (DeepModeling, ByteDance, Shanghai AI Lab, ~30 active academic groups). Japan has world-class infrastructure (Matlantis, phonopy, RIKEN AIP) but few foundation-MLIP architecture papers. Korea (Han) and Singapore (Hippalgaonkar/Canepa/IHPC) are growing fastest in per-capita output. Hong Kong is genuinely under-staffed in MLIPs and represents a 2–3-year hiring window.

Notable funding programs and consortia

  • EU: ERC Starting/Consolidator/Advanced (Csányi, Deringer, Maurer, von Lilienfeld, Cheng, Noé all hold these); Horizon Europe MaX Centre of Excellence; ELLIS network; NCCR MARVEL (Switzerland); BIFOLD (Germany); SIMPLAIX consortium (Heidelberg).
  • UK: UKRI Future Leaders Fellowships, EPSRC New Investigator Awards.
  • Germany: DFG Heisenberg, SFBs, Humboldt Professorships (Maurer 2026 €5M).
  • US: NSF AI Institutes (multiple), NSF CAREER, DOE Early Career Awards, AFOSR YIP, Schmidt Sciences AI2050, Sloan/Dreyfus Foundation, MIT-IBM Watson AI Lab.
  • China: NSFC General + Distinguished Young Scientists; MOST AI+Science National Key R&D Program (the "2025ZD06xx" series); Beijing/Shanghai municipal AI institutes (AISI, Shanghai AI Lab); CAS Strategic Priority Programs; STI 2030.
  • Japan: JST PRESTO/CREST/Mirai; JSPS KAKENHI; Kyoto ESISM; NIMS ICYS; RIKEN SPDR.
  • Korea: NRF Brain Pool / Brain Pool Plus; KAIST/SNU/POSTECH internal programs; KIAS Research Fellow.
  • Singapore/HK/Australia: NRF Singapore Fellowship/CRP; Hong Kong RGC GRF/CRF/Croucher; ARC Future Fellowships, ARC DECRA, ARC Centres of Excellence (e.g. Exciton Science).
  • Industry consortia: Open Catalyst Project (Meta + CMU), OMol25/UMA author consortium (Meta + Csányi/Krishnapriyan/Eastman/Kitchin/Persson/Blau), DeepModeling consortium (DP Tech + IAPCM + AISI + PKU + Tsinghua + USTC), NVIDIA Inception (Orbital, CuspAI, Iambic, Genesis, Inductive).

Closing observation

The current academic MLIP ecosystem is bifurcating into "foundation-model" labs (Csányi, Kozinsky, Smidt, Ceriotti, Meta, MSR) and "physics-baseline" labs (Grimme, Isayev, Meuwly, DiStasio, Müller/Tkatchenko, Behler). Researchers spanning both camps — e.g. Isayev’s AIQM3 sitting on top of semi-empirical priors while training on CCSD(T)-quality data, or Cheng’s LES adding long-range electrostatics atop short-range equivariant cores — are the ones whose methods are likeliest to land in production deployment within 24 months. The Asia-Pacific landscape, especially Hong Kong, lacks a flagship MLIP-architecture group entirely; the next two years likely decide which institutions seed that capacity.