A strategic roadmap for an atomistic machine-learning ecosystem
Jörg Behler, Michele Ceriotti, Cecilia Clementi, Gábor Csányi, Alin-Marin Elena, Aditi Krishnapriyan, Joseph W. Abbott, Fabio Affinito, Albert P. Bartók, Ilyes Batatia, Filippo Bigi, Florian N. Brünig, Yannick Calvino Alonso, Giuseppe Carleo, Aurélie Champagne, Stefan Chmiela, Marc L. Descoteaux, Ralf Drautz, Alexandra Farcas, Meng Gao, Rohit Goswami, Michael F. Herbst, Christian Holm, James R. Kermode, Alexander L. M. Knoll, Tobias Kreiman, Hoang-Thien Luu, Yury Lysogorskiy, Mihai-Cosmin Marinica, Rocco Meli, Klaus-Robert Müller, Frank Noé, Mohamadhosein Nosratjoo, Simon Olsson, Christoph Ortner, Aldo S. Pasos-Trejo, Anyang Peng, Eric Qu, Andrea Rizzi, Mariana Rossi, Bassem Sboui, Gregor N. C. Simm, Alexandre Tkatchenko, Jacopo Venturin, O. Anatole von Lilienfeld, William C. Witt, Brandon M. Wood, Tigany Zarrouk, Fabian Zills
physics.chem-ph, cond-mat.mtrl-sci, cs.LG
2026-09-30
49 authors argue atomistic ML potentials are mature; the real bottlenecks are incompatible datasets, benchmark leakage and fragmented software stacks.
Simulating a glass of water, an alloy, or a protein hits the same wall: solving the quantum equations for the electrons is too expensive for the system sizes and timescales that matter. The fix of the past two decades is to generate energies and forces for many atomic configurations with DFT (density functional theory, the workhorse first-principles method), then train a model to reproduce that potential-energy surface and run simulations with it instead. These models, MLIPs, are the field's most mature product: specialized potentials approach electronic-structure accuracy with a few thousand training structures, and general-purpose "foundation" potentials hand researchers a usable first model of a new system with no in-house training at all.
This is a strategic roadmap from 49 researchers after a January 2026 workshop at CECAM in Lausanne, with EPFL, Cambridge, Meta FAIR, Microsoft Research and several Max Planck institutes on the author list, along with the groups behind the field's two founding techniques, HDNNP and GAP. What it diagnoses is an ecosystem disease: datasets built under different but individually reasonable DFT settings are numerically incompatible, and merging them silently passes the bias on to models; new models appear almost weekly while benchmarks fall behind; every architecture ships its own training code and data format. The central judgment is blunt: further progress now depends on training-data quality as much as on ML methodology itself.
A consensus paper runs no new experiments; the deliverable is an analysis framework plus six actions. The framework splits a model along three axes: how much physics is built in, what architecture expresses it, and what data trains it. Two conclusions deserve to be remembered:
The six actions (Sec. VII): coordinate, do not consolidate, agreeing only on language-agnostic interfaces (a calculator API, a dataset field ontology, model-card metadata) and a few shared kernels; prioritize higher-fidelity data where DFT is least reliable (strong correlation, transition metals, magnetism), with a few hundred community reference structures that every dataset recomputes at its own level of theory so incompatibilities surface before training; add biennial blind challenges alongside benchmarks, hosted by CECAM and organized by the previous winner, who cannot compete; use controlled problems with known ground truth to settle when a model extrapolates; make uncertainty quantification a default that ships with every model; fund real career paths for research software work.
A roadmap runs no comparative experiments; every number below is the paper characterizing prior work:
| Item | Number | Compared to |
| Minima/saddle search with MLIP pre-optimization plus ab initio refinement | up to 90% cheaper | pure ab initio search |
| Direct-force inference speed | 2-3x faster | conservative gradient-based forces |
| Dataset size that stays competitive when carefully curated | under 1M structures | corpora of hundreds of millions |
| Model vs. reference-method uncertainty (paper's example) | 5 vs. 50 meV/atom | same property |
Two quantitative judgments sit outside the table: on sophisticated simulation tasks, the zero-shot error of universal models against their reference method is often lower than that reference's error against experiment, which makes fine-tuning superfluous in many settings; and protein-scale machine-learned coarse-grained potentials accelerate sampling by several orders of magnitude.
For AI4Science practitioners, the paper is a map with the open lots marked: electronic-structure surrogates that predict Hamiltonians and density matrices rather than energy surfaces, learnable coarse-graining maps, learned integrators, and neural DFT functionals positioned as data generators for the next generation of potentials.
For the broader ML audience, the field is re-treading LLM-era ground: benchmarks decay once they become targets, test structures migrate into later training sets exactly as LLM evaluations were contaminated, and the Clever Hans effect (right answers for wrong reasons) gets an explicit warning. The "quantum ceiling" is a concept worth borrowing: a surrogate's accuracy against experiment is capped by its teacher, so a model with zero generalization error still reproduces the DFT functional's self-interaction error. That logic extends to any surrogate trained on an imperfect reference.
One more angle you rarely see in a physics roadmap: as coding agents drive the cost of glue code and of chasing upstream breaking changes toward zero, the moat around monolithic scientific codebases shrinks and the economics of modularity reverse. A workshop reasoning carefully about how AI agents change software-architecture economics is itself a signal.
The paper concedes much of this itself: none of the six actions is technically novel, everything rides on community adoption, and governance is sketched only as a rotating committee; blind challenges depend on fresh experimental measurements and are resource-intensive, with LAM Crystal Philately cited as the sole emerging example.
Points that read as weak: the equivariance debate is presented with both sides and no verdict, so a newcomer still cannot choose an architecture from this document; the reference-structure fingerprint plan leaves open who computes it, who hosts it, and what happens to datasets that fail; and every headline number is transcribed from cited work rather than measured here. The scope is also narrower than the title suggests: AlphaFold-style sequence models, cheminformatics, and post-processing pipelines are explicitly excluded.