Multistage gap fit Turbo_SOAP

[Felixrccs]

I am starting to use the multistage gap fit with the turbo-soap descriptor. bapatist and I adjusted the multistage_gap_params.json file that you provide in the examples/iterative_gap_fit folder to be suited for the turbo soap-descriptor (in combination with the universal SOAP package). Since the turbo soap descriptor is currently quite popular with the GAP users, I think it would be wise to add our turbosoap_multistage.json file either to the examples or in the Documentation. As soon as I get your feedback I can do the regarding pull request.
{ "stages": [ { "error_scale_factor": 10.0, "descriptors": [ { "descriptor": { "distance_Nb": true, "order": 2, "cutoff": "_EVAL_ {BOND_LEN_Z1_Z2}*1.5", "cutoff_transition_width": "_EVAL_ {BOND_LEN_Z1_Z2}*1.5/5.0", "compact_clusters": true, "Z": ["_EVAL_ {Z1}", "_EVAL_ {Z2}"] }, "fit": { "n_sparse": 20, "covariance_type": "ard_se", "theta_uniform": "_EVAL_ {BOND_LEN_Z1_Z2}*1.5/5.0", "sparse_method": "uniform", "f0": 0.0 }, "count_cutoff": "_EVAL_ {BOND_LEN_Z1_Z2}*1.4", "add_species": "manual_Z_pair" } ] }, { "error_scale_factor": 1.0, "descriptors": [ { "descriptor": { "soap_turbo": true, "alpha_max": "_EVAL_ {nZ}*[8]", "l_max": 4, "atom_sigma_r": "_EVAL_ {nZ}*[{ATOM_SIGMA}]", "atom_sigma_t": "_EVAL_ {nZ}*[{ATOM_SIGMA}]", "atom_sigma_r_scaling": "_EVAL_ {nZ}*[0.0]", "atom_sigma_t_scaling": "_EVAL_ {nZ}*[0.0]", "amplitude_scaling": "_EVAL_ {nZ}*[1.0]", "rcut_hard": "_EVAL_ {R_CUT}", "rcut_soft": "_EVAL_ {R_CUT} - {R_TRANS}", "central_weight": "_EVAL_ {nZ}*[1.0]", "n_species": "_EVAL_ {nZ}", "species_Z": "_EVAL_ {Zs}", "radial_enhancement": 1, "basis": "poly3gauss", "Z": "_EVAL_ {Zcenter}" }, "add_species": "manual_universal_SOAP", "fit": { "n_sparse": 1000, "f0": 0.0, "covariance_type": "dot_product", "zeta": 4.0, "sparse_method": "cur_points", "print_sparse_index": true } } ] } ], "gap_params": { "default_sigma": [0.0025, 0.0150, 0.125, 0.125], "sparse_jitter": 1e-08, "do_copy_at_file": false, "sparse_separate_file": true, "energy_parameter_name": "DFT_energy", "force_parameter_name": "DFT_forces" } }

[bernstei]

I agree that the parameters are different enough that it's worth having an example. I'm also using turboSOAP with the multi-stage universal_SOAP heuristic GAP, so let me compare yours with mine, and confirm that everything looks right.

[Felixrccs]

turbosoap_multistage.zip

Here is the full json file. that is probably easier to look at.