Prediction of overlap matrix.

[flamingoXu]

Dear Professor Zhong,

I would like to ask if it is possible to set the 'property' parameter to both 'overlap' and 'hamiltonian' simultaneously in order to obtain both the S matrix and H matrix together. From my understanding, the file overlap.scfout generated by postprocess_openmx contains an accurate result, even though it is not directly from an SCF calculation.

To obtain the overlap matrix, should I enable the 'overlap' option during the training process?

Thank you for your guidance.

[QuantumLab-ZY]

Dear @flamingoXu,

To predict the overlap matrix as well, set the ham_only parameter to false. Additionally, you need to add the overlap key in both the losses and metrics sections of your configuration. However, I believe it's unnecessary to predict the overlap matrix. Predicted overlap matrices can sometimes lead to issues, such as being non-positive definite when diagonalizing the Hamiltonian matrix.

Best regards, Yang Zhong

[flamingoXu]

Dear Professor Zhong,

Thank you for your response. The overlap matrix is essential for calculating nonadiabatic couplings in nonadiabatic molecular dynamics (NA-MD), so I am exploring the use of a model to assist with this. I was wondering if it is possible to retrain an existing model—one that has already considered the Hamiltonian and band energy—to also account for the overlap matrix?

Additionally, could you provide more details on the issue of a non-positive definite overlap matrix? If this occurs, what strategies can be used to address or mitigate its impact?

Thank you for your time and assistance.

xuhui xu

[QuantumLab-ZY]

Dear @flamingoXu,

We've previously used HamGNN for NAMD calculations. The overlap matrix doesn't need prediction because it can be calculated analytically. The issue with non-positive definite overlap matrices is complex and mathematical.

I've attempted using a predicted overlap matrix for Hamiltonian diagonalization, and even when the predicted overlap matrix appears highly accurate—often more so than the Hamiltonian—it still can cause non-positive definite errors during diagonalization. This is especially problematic in large systems where minor deviations can lead to failures.

Since the overlap matrix is predicted, it's impossible to eliminate all potential errors. Even OpenMX might occasionally report a non-positive definite error, though this is rare.

Best regards, Yang Zhong