search

google-deepmind / deepmind-research

This repository contains implementations and illustrative code to accompany DeepMind publications
Apache License 2.0
13.08k stars 2.57k forks source link

SOSCF with dm21 in pySCF #479

Open HengZ-921 opened 10 months ago

HengZ-921 commented 10 months ago

Hello, I have tried to apply dm21 to transitional metal systems recently. But even with single atom like V, Cr, Mn, I found they are very hard to converge. I tried to apply SOSCF by mf = scf.UKS(mol).density_fit() mf._numint = dm21.NeuralNumInt(dm21.Functional.DM21) mf = mf.newton() mf.kernel(mo_init,mocc_init) which is basically the example code for SOSCF in pySCF example.

but it gave me the error message

File "/cluster/home/vuckovics/miniconda3/envs/pyscf-2.4.0/lib/python3.11/site-packages/pyscf/soscf/newton_ah.py", line 853, in kernel return _CIAH_SOSCF.kernel(self, mo_coeff, mo_occ, dm0) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/cluster/home/vuckovics/miniconda3/envs/pyscf-2.4.0/lib/python3.11/site-packages/pyscf/soscf/newton_ah.py", line 738, in kernel kernel(self, mo_coeff, mo_occ, dm0, conv_tol=self.conv_tol, File "/cluster/home/vuckovics/miniconda3/envs/pyscf-2.4.0/lib/python3.11/site-packages/pyscf/soscf/newton_ah.py", line 533, in kernel rotaiter.send((mo_coeff, mo_occ, dm, vhf, e_tot)) File "/cluster/home/vuckovics/miniconda3/envs/pyscf-2.4.0/lib/python3.11/site-packages/pyscf/soscf/newton_ah.py", line 339, in _rotate_orb_cc g_kf, h_op, h_diag = mf.gen_g_hop(mo_coeff, mo_occ, fock_ao) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/cluster/home/vuckovics/miniconda3/envs/pyscf-2.4.0/lib/python3.11/site-packages/pyscf/soscf/newton_ah.py", line 205, in gen_g_hop_uhf vind = mf.gen_response(mo_coeff, mo_occ, hermi=1) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/cluster/home/vuckovics/miniconda3/envs/pyscf-2.4.0/lib/python3.11/site-packages/pyscf/scf/_response_functions.py", line 171, in _gen_uhf_response rho0, vxc, fxc = ni.cache_xc_kernel(mol, mf.grids, mf.xc, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/cluster/home/vuckovics/miniconda3/envs/pyscf-2.4.0/lib/python3.11/site-packages/pyscf/dft/numint.py", line 2578, in cache_xc_kernel vxc, fxc = ni.eval_xc_eff(xc_code, rho, deriv=2, xctype=xctype)[1:3] ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/cluster/home/vuckovics/miniconda3/envs/pyscf-2.4.0/lib/python3.11/site-packages/pyscf/dft/numint.py", line 2711, in eval_xc_eff exc, vxc, fxc, kxc = self.eval_xc(xc_code, rhop, spin, 0, deriv, omega, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/cluster/home/vuckovics/miniconda3/envs/pyscf-2.4.0/lib/python3.11/site-packages/density_functional_approximation_dm21/neural_numint.py", line 711, in eval_xc mol=self._system_state.mol, ^^^^^^^^^^^^^^^^^^^^^^ AttributeError: 'NoneType' object has no attribute 'mol'

Could you please tell me how to apply newton properly with dm21?

jsspencer commented 9 months ago

Sorry, DM21 doesn't support second-order derivatives required for Newton optimization (and the way NumInt is patched in the PySCF-DM21 interface would require some careful handling even if it did, as your error shows).

A quasi-Newton method can be used (e.g. one from scipy). You might also find it useful to start from an initial guess produced by a cheaper functional (e.g. B3LYP, as shown in the README)

DeepDevIA commented 9 months ago

Sorry, DM21 doesn't support second-order derivatives required for Newton optimization (and the way NumInt is patched in the PySCF-DM21 interface would require some careful handling even if it did, as your error shows).

A quasi-Newton method can be used (e.g. one from scipy). You might also find it useful to start from an initial guess produced by a cheaper functional (e.g. B3LYP, as shown in the README)

Hi jsspencer how about Transition state ..!,I tried to reproduce the Bicyclobutane example (Fig 3 C of the paper) and I got the same error message: AttributeError: 'NoneType' object has no attribute 'mol' in there any way to do this please!! thanks

jsspencer commented 9 months ago

Did you follow the steps in https://github.com/google-deepmind/deepmind-research/tree/master/density_functional_approximation_dm21#best-practices-for-using-the-neural-functionals (but for a single molecule)? I don't know why you would trigger that AttributeError -- hard to know without code.

DeepDevIA commented 9 months ago

maybe I wrote my code wrong.!? I will review my code carefully. thanks for observing.