Questions about MACE's fine-tuning and active learning

[MasterLucas]

Dear MACE developers,

I have two questions regarding MACE's fine-tuning options and active learning.

1. When fine-tuning a foundation model — either using the naïve or multi-head methods — does MACE freeze any layers by default, or are all model weights updated? Is it possible to select specific layers to freeze?
2. Regarding the active learning scheme, could you clarify the origin or derivation of the uncertainty expression used in the following code snippet?

def stop_error(dyn, threshold, reg=0.2):
    atomsi = dyn.atoms
    force_var = np.var(atomsi.calc.results["forces_comm"], axis=0)
    force = atomsi.get_forces()
    ferr = np.sqrt(np.sum(force_var, axis=1))
    ferr_rel = ferr / (np.linalg.norm(force, axis=1) + reg)

    if np.max(ferr_rel) > threshold:
        print(
            "Error too large {:.3}. Stopping t={:.2} fs.".format(  # pylint: disable=C0209
                np.max(ferr_rel), dyn.get_time() / units.fs
            ),
            flush=True,
        )
        dyn.max_steps = 0

I found this snippet in the active_learning_md.py file in the cli directory.

Yours faithfully,

Lucas Bandeira

[ilyes319]

Hey @MasterLucas,

• At the moment, MACE does not freeze any weight when doing the finetuning. There is a PR to add some option to do freezing finetuning here: mace-freeze merge #830 (if you are interested, please look at the paper too). This will be merged soon in the main branch as part of a reworking of our finetuning interface. We will add new options like LORA finetuning too.

• For the uncertainty, I think @LarsSchaaf implemented it. I is similar to the Hyperactive learning expression from here: https://www.nature.com/articles/s41524-023-01104-6