PAO-ML stands for Polarized Atomic Orbitals from Machine Learning. It uses machine learning to generate geometry adopted small basis sets. It also provides exact ionic forces. The scheme can serve as an almost drop-in replacement for conventional basis sets to speedup otherwise standard DFT calculations. The method is similar to semi-empirical models based on minimal basis sets, but offers improved accuracy and quasi-automatic parameterization. However, the method is still in an early stage - so use with caution. For more information see: 10.3929/ethz-a-010819495.
The PAO-ML scheme takes a set of training structures as input. For each of these structures, the variational PAO basis is determined via an explicit optimization. The training structures should be much smaller than the target system, but large enough to contain all the motifs of the larger system. For liquids a good way to obtain structures is to run an MD of a smaller box.
Choose a primary basis set, e.g.
DZVP-MOLOPT-GTH and perform a full LS_SCF optimization. You should also enable RESTART_WRITE to save the final density matrix. It can be used to speed up the next step significantly.
Choose a PAO_BASIS_SIZE for each atomic kind. Good results can already be optained with a minimal basis sets. Slightly larger-than-minimal PAO basis sets can significantly increase the accuracy. However, they are also tougher to optimize and machine learn.
Most of the PAO settings are in the PAO sections:
&PAO EPS_PAO 1.0E-7 ! convergence threshold of PAO optimization MAX_PAO 10000 ! minimal PAO basis usually converge withing 2000 steps. MAX_CYCLES 500 ! tunning parameter for PAO optimization scheme MIXING 0.5 ! tunning parameter for PAO optimization scheme PREOPT_DM_FILE primay_basis.dm ! restart DM from primary basis for great speedup LINPOT_REGULARIZATION_DELTA 1E-6 !!!! Critical parameter for accuracy vs learnability trade-off !!!! LINPOT_REGULARIZATION_STRENGTH 1E-3 ! rather insensitive parameter, 1e-3 works usually REGULARIZATION 1.0E-3 ! rather insensitive parameter, 1e-3 works usually PRECONDITION YES ! not important, don't touch LINPOT_PRECONDITION_DELTA 0.01 ! not important, don't touch LINPOT_INITGUESS_DELTA 1E+10 ! not important, don't touch &PRINT &RESTART BACKUP_COPIES 1 ! write restart files, just in case &END RESTART &END PRINT &END PAO
Settings for individual atomic kinds are in the KIND section:
&KIND H PAO_BASIS_SIZE 1 ! set this to at least the minimal basis size &PAO_POTENTIAL MAXL 4 ! 4 works usually BETA 2.0 ! 2 work usually, but is worth exploring in case of accuracy or learnability issues. &END PAO_POTENTIAL &END KIND
Finding the optimal PAO basis poses an intricate minimization problem, because the rotation matrix U and the Kohn-Sham matrix H have to be optimized in a self-consistent manner. In order to speedup the optimization, the Kohn-Sham matrix is only updated occasionally while most time is spend on optimizing U. This alternating scheme is controlled by two input parameters:
The progress of the PAO optimization can be tracked from lines that start with
PAO| step. The columns have the following meaning:
step-num energy conv-crit. step-length time PAO| step 1121 -186.164843303 0.227E-06 0.120E+01 1.440
For the simulation of larger systems the PAO-ML scheme infers new PAO basis sets from the training data. For this two heuristics are employed: A descriptor and an inference algorithm. Currently, only one simple descriptor and Gaussian processes are implemented. However, this part offers great opportunities for future research.
In order to obtain good results from the learning machinery a small number of so-called hyperparameters have to be carefully tuned for each application. For the current implementation this includes the GP_SCALE and the descriptor's BETA and SCREENING.
For the optimization of the hyper-parameter exists no gradient, hence one has to use a derivative-free method like the one by Powell. A versatile implementation is e.g. the scriptmini tool. A good optimization criterion is the variance of the energy difference wrt. the primary basis across the training set. Alternatively, atomic forces could be compared. Despite the missing gradients, this optimization is rather quick because it only performs calculations in the small PAO basis set.
Most of the PAO-ML settings are in the PAO/MACHINE_LEARNING sections:
&PAO MAX_PAO 0 ! use PAO basis as predicted by ML, required for correct forces PENALTY_STRENGTH 0.0 ! disable penalty, required for correct forces &MACHINE_LEARNING GP_SCALE 0.46 !!! critical tuning parameter - depends also on descriptor settings !!! GP_NOISE_VAR 0.0001 ! insensitive parameter METHOD GAUSSIAN_PROCESS ! only implemented method - opportunity for future research DESCRIPTOR OVERLAP ! only implemented method - opportunity for future research PRIOR MEAN ! try once ZERO - makes usually no difference TOLERANCE 1000.0 ! disable check for max variance of GP prediction &TRAINING_SET ../training/Frame0000/calc_pao_ref-1_0.pao ../training/Frame0100/calc_pao_ref-1_0.pao ../training/Frame0200/calc_pao_ref-1_0.pao ! add more ... &END TRAINING_SET &END MACHINE_LEARNING &END PAO
Settings for individual atomic kinds are again in the KIND section:
&KIND H PAO_BASIS_SIZE 1 ! use same settings as for training &PAO_POTENTIAL MAXL 4 ! use same settings as for training BETA 2.0 ! use same settings as for training &END PAO_POTENTIAL &PAO_DESCRIPTOR BETA 0.16 !!! important ML hyper-parameter !!! SCREENING 0.66 !!! important ML hyper-parameter !!! WEIGHT 1.0 ! usually not needed when BETA and SCREENING are choose properly &END PAO_DESCRIPTOR &END KIND
When optimizing the PAO reference data in Step 3 one has to make a trade-off between accuracy and learnability. Good learnability means that similar structures leads to similar PAO parameters. In other words the PAO parameters should depend smoothly on the atomic positions. In general, the settings presented above should yield good results. However, if problems arise in the later machine learning steps, this might be the culprit.
Unfortunately, there is not yet a simple way to assess learnability. One way to investigate is to create a set of structures along a reaction coordinate, e.g. a dimer dissociation. One can then plot the numbers from the
Xblock in the
.pao files vs. the reaction coordinate.