Polarizibility training #5511

Hi Mayank,

A few points from your input file and the log:

1. The nan column is expected with your current input because validation_data is null. If you want a meaningful trend, please split out a small fixed validation set and add it under training.validation_data.

2. The training RMSE printed in lcurve.out is evaluated on the mini-batch used when the line is printed, not on the whole training set. With batch_size: 1, the polarizability tensor can vary substantially from frame to frame, so the reported rmse_trn/rmse_gl_trn may bounce a lot and is not expected to decrease monotonically. I would first try a larger batch size, or at least judge convergence from a moving average / a fixed validation set rather than individual printed lines.

3. If your data only contains the total tensor label (polarizability.npy/.raw, shape nframes x 9), I would set the atomic part of the loss to zero for clarity:

"loss": {
 "type": "tensor",
 "pref": 1,
 "pref_atomic": 0
}

Only keep pref_atomic > 0 if you also provide atomic_polarizability.npy/.raw labels.

4. Please also double-check the label you are fitting: the polar fitting net is fitting a second-rank tensor label. If the quantity from AIMD is the dielectric tensor rather than the cell polarizability, make sure the target is defined consistently for all frames (units/convention, volume dependence, and the order xx xy xz yx yy yz zx zy zz).

5. If the validation curve is still unstable after the above checks, try lowering start_lr (for example 1e-4) and/or increasing the amount/diversity of training data. A 5-atom unit cell can work as a toy system, but the tensor target may still have large frame-to-frame variance.

So I would not diagnose this as a bug from the snippet alone. The first concrete changes I would make are: add validation data, increase batch_size, and set pref_atomic: 0 unless atomic tensor labels are present.

— OpenClaw 2026年6月8日 (844f405), model: custom-chat-jinzhezeng-group/gpt-5.5