Hello,
I was wondering if anyone had any tips on the most efficient approach to train an accurate MTP that could be used to create and relax a high-κ metal gate stack. I would like to generate an MTP similar to QuantumATK_MTP_Si_SiO2_HfO2_TiN_2022_12 (Hf, N, Si, O, Ti) but with different gate stack materials.
Is it best to build the potential additively — meaning, first train on the channel material (in my case an amorphous metal-oxide semiconductor), then add training data for the interface with the gate oxide, and finally add data for the interface with the metal gate?
In any case, what should the training data look like? Should we use the CrystalInterfaceTrainingParameters class? If so, how does this class handle training across multiple interfaces? In my example, there are two distinct interfaces that need to be accurately captured in the training. How would the process change if we want the MTP to be able to handle more interfaces (3 or more)? Like for example if we wanted to be able to handle interfacial oxide (SiO2) layers.
Also, how can we add realistic interface structures to the training set without running full ab initio MD simulations? I ask because my understanding is that the training data generated by CrystalInterfaceTrainingParameters would primarily consist of cleaved and matched bulk structures, which might not represent realistic relaxed interfaces.
I would really appreciate any advice or shared experiences.
Thank you!
