[Docs] Add more FAQs about the chocie of parameters

[njzjz]

Summary

Currently, we have a benchmark of network size in the FAQ. I think we should choose some typical datasets to test more parameters and benchmark their accuracy and speed (for both CPUs and GPUs).

Detailed Description

I have several ideas here:

• [ ] type_one_side: #2265
• [ ] precision: In some (or most) systems, FP64 does not improve accuracy compared to FP32.
• [ ] activation_function
• [ ] the type of the descriptor: se_e2_a, se_e2_r, se_e3, local_frame
• [ ] axis_neuron
• [ ] training steps
• [ ] loss function, i.e., the weight of energy/forces
• [ ] learning rate
• [ ] batch size

Further Information, Files, and Links

No response

[wangsl613]

Given that FP64 does not enhance accuracy over FP32, how can I switch from FP64 to FP32 in deepmd during training? Please let me know. Many thanks!