Frequently Asked Questions¶

What kind of properties can I fit?¶

In AMS2024 you can only train to

Yes, both the training and production simulations can be run on a compatible GPU when training/running M3GNet.

Running on the GPU results in significant speedup so it is recommended to do so.

To use the GPU you must install a GPU-compatible version of M3GNet or other machine learning backends.

See the MLPotential and ParAMS documentations.

Note

The engines typically used for DFT reference calculations, like ADF, BAND, and the AMS-bundled version of Quantum ESPRESSO, cannot run on the GPU.

The Active Learning procedure repeatedly switches between ML training, ML production simulations, and reference calculations.

Simple Active Learning is compatible with (almost) all types of MD simulations supported by the AMS Driver.

	Type	Comment
✔	NVE
✔	NVT	example \| tutorial
✔	NPT
✔	Temperature ramping	example
✔	Multiple thermostats (thermal conductivity)
✔	Molecule gun or sink (CVD, ALD)
✔	CREST metadynamics	example
✔	Constraints (e.g., fixed positions)
✔	Apply force or velocity (friction and viscosity)
✔	Lattice deformation linear	Use Steps%Type = Linear \| example
✔	Reaction boost force	Use committee uncertainties \| example
✔	Reaction boost target coordinate	Use Steps%Type = Linear \| example
✔	Force-bias Monte Carlo (fbMC)
(✔)	Lattice deformation periodic	Difficult setup to get good training/validation sets
(✔)	Nanoreactor	Difficult setup to get good training/validation data
❌	Very high temperatures (combustion)	The workflow automatically considers very high temperatures (> 5000 K) to be unrealistic and assumes that they are artifacts of the model