Behler roadmap: "Coordinate, do not consolidate" atomistic ML

The roadmap grew out of a January 2026 CECAM workshop in Lausanne, where researchers discussed algorithms, models, software and hardware infrastructure, and promising applications. Its long list of co-authors — Jörg Behler, Michele Ceriotti, Cecilia Clementi, Gábor Csányi and dozens more — reflects the point that building truly general models is not work a few groups can carry alone.

The field moved from the 2007 Behler–Parrinello potential, which assembled a whole system's energy from local atomic environments, through rotation-aware methods such as NequIP that achieve high accuracy with less training data, to today's pretrained general models like MACE-MP, UMA, DPA-2, MACE-Osaka24 and SevenNet-Omni. Those models differ in how they handle data computed under different DFT settings: some align energies beforehand, others feed the calculation conditions into the model and condition on them.

A recurring theme is that prediction accuracy is not the same as simulation value. A model can describe local energies and gradients accurately yet predict polymer density far from experiment, and an ACE comparison of ensemble methods and D-optimality found error underestimated on heterogeneous datasets. That is why the roadmap treats uncertainty calibration, blind testing and reproducible evaluation as shared infrastructure rather than afterthoughts.