Incorrect results will be given in these versions and situations with no error messages:
v3.0.0~v3.2.0b0: in the PyTorch backend with LAMMPS, the compressed se_e2_a model may produced incorrect results; reported by the MD running is only working when using CPU #5092 fix(pt): sort nlist for compressed se_e2_a in forward_lower #5524
v3.0.0~v3.0.2: when using DeepEval to evaluate DPLR models, long-range term is not added; reported by [BUG] Incorrect results from DPLR model in python inference interface #4625 fix(tf): fix dplr Python inference #4753
v2.2.5~v2.2.7: random incorrect GPU results due to invalid GPU memory write (see [BUG] _Strange problems when running MD with a DPA-1 model_ #3150 [BUG] _Illegal memory access error in LAMMPS caused by the coupling of fix nvt command and compute command_ #3102 [BUG] DP compress+DPLR+GPU results in bug in building neighbor list in LAMMPS #3094 Some problems encountered in using GPU to accelerate lammps #3034 [BUG] Model deviation calculation output nan occasionally in model_devi.out #3242 Fix max nbor size related issues #3157
~v2.2.6: the compressed model generates incorrect results when neuron[0] is 1; fixed by support compressing any neuron structure #2933
~v2.2.6: the compressed se_e2_a model generates incorrect results for symmetrical coordinates when sel is just enough (this condition is rare as most structures during simulations are not symmetrical); fixed by fix se_a compression for just enough sel and symmetrical coordinates #2924
v2.2.4~v2.2.5: se_atten_v2 (or se_atten with smooth_type_embdding=True) model gives inconsistent energy and force; fixed by fix: make the se attn v2 descriptor energy conservative. #2905
v2.2.3~v2.2.5: incorrect results when compressed training se_atten model; reported by [BUG] Significant errors when compressed training se_atten_v2 model #2865 Se atten grad grad #2898
v2.2.0~v2.2.4: incorrect results from DPLR training; reported by [BUG] Different results from different versions when training DPLR model #2862 fix the issue of applying modifier multiple times when batch set is load only once #2864
~v2.2.3 with TF >=2.13: incorrect results on GPUs; reported by [BUG] [critical] TF v2.13.0 calculates wrong GPU results #2660 synchronize in the beginning of all CUDA functions #2661
v2.1.0~v2.2.1: incorrect results from se_e3 compressed model; reported by A large error induced by compression for se_e3 descriptor #2250 fix se_e3 tabulate op #2552
v2.2.0: frozen models can not be performed on LAMMPS; reported by INVALID_ARGUMENT: Incompatible shapes: [1,96] vs. [1,4020] [[{{node gradients/mul_4_grad/BroadcastGradientArgs}}]] ERROR: DeePMD-kit Error: TensorFlow Error: INVALID_ARGUMENT: Incompatible shapes: [1,96] vs. [1,4020] [[{node gradients/mul_4_grad/BroadcastGradientArgs}]] (/home/conda/feedstock_root/build_artifacts/libdeepmd_1678813995417/work/source/lmp/pair_deepmd.cpp:471) #2398 atype_filter should have the shape of nloc #2390
v2.1.0~v2.1.4: incorrect force results on CPUs when rcut * 2 > box size; fixed by revert prod_force OMP in #1360
#1862
~v2.1.0 with LAMMPS >=30Jul2021: incorrect LAMMPS pppm_dplr results; fixed by Correct the forward communication at ik differentiation mode in pppm_dplr #1637
~v2.0.3 with LAMMPS >=30Jul2021: incorrect LAMMPS compute deeptensor/atom results; reported by [BUG] Incorrect dipole prediction with LAMMPS of new version #1381 explicitly set neighbor request to full in compute deeptensor/atom to fix bug #1381 #1382
~v2.0.1 with LAMMPS >=30Jul2021: incorrect LAMMPS pair deepmd results; reported by [BUG] In the release 2.0.0 version, the lammps result is abnormal #1109 explicitly set neighbor request to full to fix bug #1109 #1128
The potential energy is not continuous in the following cases:
Incorrect results will be given in these versions and situations with no error messages:
se_e2_amodel may produced incorrect results; reported by the MD running is only working when using CPU #5092, fixed by fix(pt): sort nlist for compressed se_e2_a in forward_lower #5524neuron[0]is 1; fixed by support compressing any neuron structure #2933se_e2_amodel generates incorrect results for symmetrical coordinates whenselis just enough (this condition is rare as most structures during simulations are not symmetrical); fixed by fix se_a compression for just enough sel and symmetrical coordinates #2924se_atten_v2(orse_attenwithsmooth_type_embdding=True) model gives inconsistent energy and force; fixed by fix: make the se attn v2 descriptor energy conservative. #2905se_attenmodel; reported by [BUG] Significant errors when compressed training se_atten_v2 model #2865, fixed by Se atten grad grad #2898se_e3compressed model; reported by A large error induced by compression for se_e3 descriptor #2250, fixed by fix se_e3 tabulate op #2552pppm_dplrresults; fixed by Correct the forward communication at ik differentiation mode in pppm_dplr #1637compute deeptensor/atomresults; reported by [BUG] Incorrect dipole prediction with LAMMPS of new version #1381, fixed by explicitly set neighbor request to full in compute deeptensor/atom to fix bug #1381 #1382pair deepmdresults; reported by [BUG] In the release 2.0.0 version, the lammps result is abnormal #1109, fixed by explicitly set neighbor request to full to fix bug #1109 #1128The potential energy is not continuous in the following cases:
loc_framemodelse_attenmodel withset_davg_zero=False(default=False in v2.1.5 and True in >=v2.2.0) andsmooth_type_embdding=False(added in v2.2.4, default=False)se_attenorse_atten_v2model with non-zeroattn_layer; reported by [BUG] Error while training DPLR #2842, fixed by fix: make the se attn v2 descriptor energy conservative. #2905