Introduction
Tadah! is a modular C++17 framework for developing and deploying machine learning interatomic potentials (MLIPs). It ships:
a command-line driver,
tadah, for training, prediction, dataset manipulation, and hyperparameter optimisation (tadah hpo); anda LAMMPS pair style,
pair_style tadah, that runs trained potentials in MD without a Python or Tadah! runtime dependency.
The two pieces correspond to two repositories — Tadah!MLIP (training
side) and Tadah!LAMMPS (deployment side) — that share a common pot.tadah
file format.
Key features
Pluggable descriptors and regressors — mix descriptor families (2-body, many-body, EAM-style) with linear or kernel regressors.
Nested fitting (HPO) — drive the training loop from an outer optimiser that scores trial potentials against physics-informed constraints (elastic constants, equilibrium volume, surface energies, …). See Nested Fitting.
LAMMPS interface — every descriptor and model exposed through
pair_style tadah;LSCALEandESHIFTround-trip through the potential file.OpenMP build — desktop parallelism via
OMP_NUM_THREADS. An MPI build target exists but is not functional in the 1.3.0-beta.1 release (see Installation).C++ API — link against
libtadah.mlip/libtadah.corefor embedded use; see Compiling and Linking with Library.
CLI surface
The tadah driver exposes the following top-level commands (each
documented in Command Line Interface):
tadah train/tadah predict— fit and apply potentials.tadah hpo— nested fitting (hyperparameter optimisation).tadah data— convert,balance,dedup,merge,sample,split,print,writedatasets and structures (CIF, VASP, CASTEP, LAMMPS; online sources MP/COD/NOMAD via--structure).tadah analysis— plot basis functions, cutoffs, descriptors.tadah properties pairwise— pairwise energy between two atoms.tadah explain <key>— print the help text for any configuration key.
Obtaining Tadah!
Tadah!MLIP and Tadah!LAMMPS are hosted at:
For build instructions see Installation. Selected published potentials produced with Tadah! by Prof. Ackland's group are listed in Trained MLIPs.