run LAMMPS problem

[KK33999]

Hello everyone, I have a question to discuss. I trained the potential function using the DeePMD-kit v3.0.2 version DPA2 descriptor, and the energy and force accuracy were also in good agreement during the dp test process. However, when I used this potential function to run LAMMPS molecular dynamics, I found that it would scatter the structure after 20 ps. The specific details of my training are: 1. Select the temperature range of 50K-500K and use VASP to run AIMD. Check that the structure of the last step is not scattered and select 80% of all data for training. 2. Use the LAMMPS template as follows.

variable			t equal 50

log				log_$t.lammps  
units           		metal                                                  
boundary        		p p p           
atom_style     		atomic # full  # charge  dipole     
read_data	  		stru.data  
mass			1  A      
mass 			2  B

group			A  type 1
group			B  type 2     

pair_style	deepmd 	./graph.pth        
pair_coeff  		* *	                                 
velocity        		all create $t 23456789 dist gaussian

timestep        		0.001
fix             		1 all nvt temp $t $t 0.1 
                                               
thermo_style    	custom step temp pe ke etotal press density lx ly lz vol         
thermo          		100                                                                                         

dump			1 all custom 10 $t.dump id type x y z 

thermo			10                       

run             		500000

My guess is whether AIMD data needs to be cleaned? I don't know, I hope someone can give some suggestions or similar experiences.

[njzjz]

Do you know the active learning? It may be just because the data is not enough.

[KK33999]

Finally, I expanded the dataset with dpgen and achieved good results. Thank you for your answer. Best wishes.