Software
The grouping below is a maintenance signal: what we develop and support, what was released to reproduce a single paper, what is kept for the record, and what builds this site.
Actively developed
Martignac
Computational workflows for coarse-grained Martini simulations. A simulation is represented as a directed acyclic graph carrying its entire history, from system preparation to property calculation, and the framework connects to the NOMAD database so that results are stored according to the FAIR principles and an existing calculation is retrieved rather than repeated.
MIT · Python, installed from source · includes a Streamlit dashboard
Auto-Martini M3
Automatic generation of Martini 3 topologies for small organic molecules of up to 25 heavy atoms, directly from SMILES: mapping, bead types, and bonded parameters, with no human in the loop. Validated against expert-designed models and exercised on a screen of 100,000 compounds from the Enamine library. Developed and maintained with the Martini Force Field Initiative.
GPL-2.0 · Python, conda environment supplied · supersedes Auto Martini for Martini 2
Fokker-Planck score learning
Free-energy estimation from biased simulations, obtained by imposing the analytical Fokker–Planck steady-state solution on a learned score function rather than inferring it from data. Recovers potentials of mean force under periodic boundary conditions, and is most advantageous in the low-data regime.
MIT · pip install fpsl · documentation includes tutorials
Split-Flows
Backmapping treated as continuous-time measure transport between molecular resolutions. Augmenting a coarse-grained configuration with auxiliary noise dimensions makes the map to the fine-grained space one-to-one, which gives both conditional sampling of atomistic structures and a tractable route to mapping entropies – the information a given coarse-graining discards.
MIT · Python, installed from source with uv
Reference implementations
Code released with a specific paper, to reproduce it and to build on. Maintained only as far as that paper requires.
Neural thermodynamic integration
Free energies from a neural network potential that interpolates between two target Hamiltonians, so that a single learned model stands in for the ladder of intermediate states a conventional calculation has to simulate. Two repositories, both maintained by Bálint Máté (University of Geneva): the original formulation, and the extension to solvation where both end states are non-trivial.
MIT
Original code · J. Phys. Chem. Lett. 2024
Solvation code · J. Chem. Phys. 2025
Multi-level Bayesian optimization
Molecular discovery organized as a funnel across resolutions, exploring broadly where the model is cheap and exploiting where it is accurate. Includes chespex, a small package implementing the algorithm independently of the application, and a separate tutorial working through the method on two-bead molecules in a water–hexane mixture.
Earlier codes
Kept available because the papers that used them point here, not because they are maintained.
Auto Martini (Martini 2)
The original automatic Martini parametrization of small organic molecules, for Martini 2, later extended to variable bead-type resolution. Superseded by Auto-Martini M3.
IPML
Intermolecular potentials from a combination of physics-based and machine-learned components – atomic multipoles and many-body dispersion, with coefficients learned across conformations and composition – transferable across chemical space. GPL-2.0-or-later, hosted on the Max Planck Computing and Data Facility’s GitLab.
mbvdw
Classical, Voronoi-tessellation adaptation of the many-body van der Waals method, which dispenses with explicit electron densities.
peptideB
Top-down, implicit-solvent coarse-grained peptide model for secondary- and tertiary-structure formation, implemented for ESPResSo.
opt-mhm
Multiple-histogram (WHAM) free-energy analysis with optimized convergence.
Website tooling
my-scientific-profile
Retrieves the metadata of a publication record programmatically. ORCID decides which works are yours; half a dozen further providers describe them, and their answers are merged field by field with a declared trust order, so each field records where it came from. It is what builds the publication list and the research landscape here; about sets out the rest of the pipeline. MIT.