mlatom 2.1.0__tar.gz → 2.1.2__tar.gz

{MLatom-2.1.0 → MLatom-2.1.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: MLatom
-Version: 2.1.0
+Version: 2.1.2
 Summary: A Package for Atomistic Simulations with Machine Learning
 Home-page: http://mlatom.com
 Author: Pavlo O. Dral
@@ -71,9 +71,15 @@ Permutationally invariant kernel and self-correction are also supported.
   - permuted.
 
 
-# The KREG model
+# ML models
 The [ KREG (Kernel-ridge-regression using RE descriptor and the Gaussian kernel function )](http://dx.doi.org/10.1021/acs.jpclett.8b02469) model is the default ML method.
 
+## General-purpose ML models
+- AIQM1 (requires interfaces to other programs as described in http://MLatom.com/AIQM1)
+- Models available via interface to [TorchANI](https://doi.org/10.1021/acs.jcim.0c00451)
+  - ANI-1x
+  - ANI-1ccx
+  - ANI-2x
 
 # Model Validation
 [ ML model can be validated (generalization error can be estimated) in several ways: ](https://web.stanford.edu/~hastie/ElemStatLearn/)
@@ -103,13 +109,12 @@ MLatom can be also used to estimate first derivatives from an ML model. Two scen
 - derivatives for interfaced models
 
 
-# Cross Section
+# UV/vis spectra
 MLatom can significantly accelerate the calculation of cross-section with the Nuclear Ensemble Approach (NEA).
 
-In brief, this feature uses fewer QC calculation to achieve higher precision and reduce computational cost. You can find more detail on this paper (please cite it when using this feature:
+In brief, this feature uses fewer QC calculation to achieve higher precision and reduce computational cost. You can find more detail on this paper (please cite it when using this feature):
 
-> Bao-Xin Xue, Mario Barbatti*, Pavlo O. Dral*, [ Machine Learning for Absorption Cross Sections ](https://doi.org/10.1021/acs.jpca.0c05310), J. Phys. Chem. A 2020, 124, 7199–7210. DOI: 10.1021/acs.jpca.0c05310.
-> [ Preprint on ChemRxiv ](https://doi.org/10.26434/chemrxiv.12594191), DOI: 10.26434/chemrxiv.12594191.
+> Bao-Xin Xue, Mario Barbatti, Pavlo O. Dral, [ Machine Learning for Absorption Cross Sections ](https://doi.org/10.1021/acs.jpca.0c05310), J. Phys. Chem. A 2020, 124, 7199–7210. DOI: 10.1021/acs.jpca.0c05310.
 
 # Interfaces to 3<sup>rd</sup>-party software
 MLatom also provides interfaces to some third-party software where extra ML model types are natively implemented. It allows users to access other popular ML model types within MLatom's workflow. Currently available third-party model types are:
@@ -122,10 +127,10 @@ MLatom also provides interfaces to some third-party software where extra ML mode
 
 # About Program
 MLatom: a Package for Atomistic Simulations with Machine Learning    
-Version 2.1.0
+Version 2.1.2
 http://mlatom.com/                             
                                                                            
-Copyright (c) 2013-2021 Pavlo O. Dral                   
+Copyright (c) 2013-2022 Pavlo O. Dral                   
 http://dr-dral.com/                            
                                                                            
 All rights reserved. This work is licensed under the [Attribution-NonCommercial-NoDerivatives 4.0 International](http://creativecommons.org/licenses/by-nc-nd/4.0/) license. See LICENSE.CC-BY-NC-ND-4.0.  
@@ -136,7 +141,7 @@ Cite as:
 
 1. Pavlo O. Dral, J. Comput. Chem. 2019, 40, 2339-2347             
 2. Pavlo O. Dral, Fuchun Ge, Bao-Xin Xue, Yi-Fan Hou, Max Pinheiro Jr, Jianxing Huang, Mario Barbatti, Top. Curr. Chem. 2021, 379, 27              
-3. Pavlo O. Dral, Peikun Zheng, Bao-Xin Xue, Fuchun Ge, Yi-Fan Hou, Max Pinheiro Jr, MLatom: A Package for Atomistic Simulations with Machine Learning, version 2.1.0, Xiamen University, Xiamen, China, 2013-2021.               
+3. Pavlo O. Dral, Peikun Zheng, Bao-Xin Xue, Fuchun Ge, Yi-Fan Hou, Max Pinheiro Jr, MLatom: A Package for Atomistic Simulations with Machine Learning, version 2.1.2, Xiamen University, Xiamen, China, 2013-2022.               
 
 
 # License

{MLatom-2.1.0 → MLatom-2.1.2}/README.md

@@ -53,9 +53,15 @@ Permutationally invariant kernel and self-correction are also supported.
   - permuted.
 
 
-# The KREG model
+# ML models
 The [ KREG (Kernel-ridge-regression using RE descriptor and the Gaussian kernel function )](http://dx.doi.org/10.1021/acs.jpclett.8b02469) model is the default ML method.
 
+## General-purpose ML models
+- AIQM1 (requires interfaces to other programs as described in http://MLatom.com/AIQM1)
+- Models available via interface to [TorchANI](https://doi.org/10.1021/acs.jcim.0c00451)
+  - ANI-1x
+  - ANI-1ccx
+  - ANI-2x
 
 # Model Validation
 [ ML model can be validated (generalization error can be estimated) in several ways: ](https://web.stanford.edu/~hastie/ElemStatLearn/)
@@ -85,13 +91,12 @@ MLatom can be also used to estimate first derivatives from an ML model. Two scen
 - derivatives for interfaced models
 
 
-# Cross Section
+# UV/vis spectra
 MLatom can significantly accelerate the calculation of cross-section with the Nuclear Ensemble Approach (NEA).
 
-In brief, this feature uses fewer QC calculation to achieve higher precision and reduce computational cost. You can find more detail on this paper (please cite it when using this feature:
+In brief, this feature uses fewer QC calculation to achieve higher precision and reduce computational cost. You can find more detail on this paper (please cite it when using this feature):
 
-> Bao-Xin Xue, Mario Barbatti*, Pavlo O. Dral*, [ Machine Learning for Absorption Cross Sections ](https://doi.org/10.1021/acs.jpca.0c05310), J. Phys. Chem. A 2020, 124, 7199–7210. DOI: 10.1021/acs.jpca.0c05310.
-> [ Preprint on ChemRxiv ](https://doi.org/10.26434/chemrxiv.12594191), DOI: 10.26434/chemrxiv.12594191.
+> Bao-Xin Xue, Mario Barbatti, Pavlo O. Dral, [ Machine Learning for Absorption Cross Sections ](https://doi.org/10.1021/acs.jpca.0c05310), J. Phys. Chem. A 2020, 124, 7199–7210. DOI: 10.1021/acs.jpca.0c05310.
 
 # Interfaces to 3<sup>rd</sup>-party software
 MLatom also provides interfaces to some third-party software where extra ML model types are natively implemented. It allows users to access other popular ML model types within MLatom's workflow. Currently available third-party model types are:
@@ -104,10 +109,10 @@ MLatom also provides interfaces to some third-party software where extra ML mode
 
 # About Program
 MLatom: a Package for Atomistic Simulations with Machine Learning    
-Version 2.1.0
+Version 2.1.2
 http://mlatom.com/                             
                                                                            
-Copyright (c) 2013-2021 Pavlo O. Dral                   
+Copyright (c) 2013-2022 Pavlo O. Dral                   
 http://dr-dral.com/                            
                                                                            
 All rights reserved. This work is licensed under the [Attribution-NonCommercial-NoDerivatives 4.0 International](http://creativecommons.org/licenses/by-nc-nd/4.0/) license. See LICENSE.CC-BY-NC-ND-4.0.  
@@ -118,7 +123,7 @@ Cite as:
 
 1. Pavlo O. Dral, J. Comput. Chem. 2019, 40, 2339-2347             
 2. Pavlo O. Dral, Fuchun Ge, Bao-Xin Xue, Yi-Fan Hou, Max Pinheiro Jr, Jianxing Huang, Mario Barbatti, Top. Curr. Chem. 2021, 379, 27              
-3. Pavlo O. Dral, Peikun Zheng, Bao-Xin Xue, Fuchun Ge, Yi-Fan Hou, Max Pinheiro Jr, MLatom: A Package for Atomistic Simulations with Machine Learning, version 2.1.0, Xiamen University, Xiamen, China, 2013-2021.               
+3. Pavlo O. Dral, Peikun Zheng, Bao-Xin Xue, Fuchun Ge, Yi-Fan Hou, Max Pinheiro Jr, MLatom: A Package for Atomistic Simulations with Machine Learning, version 2.1.2, Xiamen University, Xiamen, China, 2013-2022.               
 
 
 # License

{MLatom-2.1.0 → MLatom-2.1.2}/setup.py

@@ -43,7 +43,7 @@ with open(os.path.join(here, 'README.md')) as f:
 
 setup(
     name = pkg_name,
-    version = "2.1.0",
+    version = "2.1.2",
     author = "Pavlo O. Dral",
     author_email = "admin@mlatom.com",
     license = 'CC BY-NC-ND 4.0',  # https://creativecommons.org/licenses/by-nc-nd/4.0/

{MLatom-2.1.0 → MLatom-2.1.2}/src/MLatom/AIQM1.py

@@ -119,10 +119,10 @@ class AIQM1Cls(object):
             models.append(model)
         self.models = models
 
-    def predict(self, device, d4, sae, optmol):
+    def predict(self, device, d4, sae):
         
         numbers, coordinates = get_xyz(self.xyzfile)
-        charges, mults, odm2_energies, odm2_grads, odm2_hesses = run_odm2_calculation(self.mndobin, numbers, coordinates, self.hesscalc, self.mndokw, optmol)
+        charges, mults, odm2_energies, odm2_grads, odm2_hesses = run_odm2_calculation(self.mndobin, numbers, coordinates, self.hesscalc, self.mndokw)
         shifts = cal_shifts(numbers, sae)
         if d4:
             d4_energies, d4_grads, d4_hesses = run_d4_calculation(self.dftd4bin, self.xyzfile, numbers, coordinates, charges, self.hesscalc)
@@ -151,10 +151,22 @@ class AIQM1Cls(object):
                     delta_hess = torchani.utils.hessian(coordinate, energies=delta_energy)
                     delta_hesses.append(delta_hess)
                     
-            delta_energy = torch.stack([e for e in delta_energies], dim=0).mean(0)
+            delta_energy = torch.stack([e for e in delta_energies], dim=0)
+
+            fmt = ' %-40s: %15.8f Hartree'
+            if len(self.models) == 8:
+                e_std = delta_energy.std(dim=0, unbiased=False)
+                print(fmt % ('Standard deviation of NN contribution', e_std), end='')
+                e_std = hartree2kcalmol(e_std)
+                print('%15.5f kcal/mol' % e_std)
+                np.savetxt('std', e_std.detach().numpy())
+            delta_energy = delta_energy.mean(0)
             delta_grad = torch.stack([g for g in delta_grads], dim=0).mean(0)
             energy = delta_energy[0] + odm2_energy + shift
             grad = delta_grad[0] + odm2_grad
+            print(fmt % ('NN contribution', delta_energy))
+            print(fmt % ('Sum of atomic self energies', shift))
+            print(fmt % ('ODM2* contribution', odm2_energy), file=sys.stdout)
             if self.hesscalc:
                 odm2_hess = torch.tensor(odm2_hesses[i]).to(device)
                 delta_hess = torch.stack([h for h in delta_hesses], dim=0).mean(0)
@@ -164,9 +176,12 @@ class AIQM1Cls(object):
                 d4_grad = torch.tensor(d4_grads[i]).to(device)
                 energy = energy + d4_energy
                 grad = grad + d4_grad
+                print(fmt % ('D4 contribution', d4_energy))
                 if self.hesscalc:
                     d4_hess = torch.tensor(d4_hesses[i]).to(device)
                     hess = hess + d4_hess
+            print(fmt % ('Total energy', energy))
+            print('')
             fenergy.write('%20.12f\n' % energy)
             fgrad.write('%d\n\n' % nat)
             for j in range(nat):
@@ -182,15 +197,14 @@ class AIQM1Cls(object):
             fhess.close()
         return charges, mults
 
-    def forward(self, optmol=None):
-        # optmol is used for geomopt
+    def forward(self):
         if self.level == 'dft':
             sae = self.dftsae
         elif self.level == 'cc':
             sae = self.ccsae
         self.define_aev()
         self.load_models()
-        charges, mults = self.predict(self.device, self.d4, sae, optmol)
+        charges, mults = self.predict(self.device, self.d4, sae)
         return charges, mults
 
 class args(object):
@@ -216,11 +230,13 @@ class args(object):
             import torch
             import torchani
             from torchani.utils import ChemicalSymbolsToInts
+            from torchani.units import hartree2kcalmol
         except: 
             stopper.stopMLatom('Please install all Python module required for TorchANI')
         globals()['torch'] = torch
         globals()['torchani'] = torchani
         globals()['ChemicalSymbolsToInts'] = ChemicalSymbolsToInts
+        globals()['hartree2kcalmol'] = hartree2kcalmol
         
         cls.mndobin  = mndobin
         cls.dftd4bin = dftd4bin
@@ -301,12 +317,12 @@ def get_xyz(xyzfile):
                 coordinates.append(coordinate)
     return numbers, coordinates
 
-def run_odm2_calculation(mndobin, numbers, coordinates, hesscalc, mndokw, optmol):
+def run_odm2_calculation(mndobin, numbers, coordinates, hesscalc, mndokw):
     charges = np.full(len(numbers), 0)
     mults = np.full(len(numbers), 1)
     energies = []; grads = []; hesses = []
     if mndokw: 
-        mndokeywords, charges, mults = read_mndokw(mndokw, optmol)
+        mndokeywords, charges, mults = read_mndokw(mndokw)
         if len(mndokeywords) != len(numbers):
             stopper.stopMLatom('Inconsistency of the nubmer of keywords and molecules')
     for i in range(len(numbers)):
@@ -338,16 +354,18 @@ def run_odm2_calculation(mndobin, numbers, coordinates, hesscalc, mndokw, optmol
         energies.append(energy)
         grads.append(grad)
         hesses.append(hess)
+
+    #energies = np.array(energies)
     return charges, mults, energies, grads, hesses
 
-def read_mndokw(mndokw, optmol):
+def read_mndokw(mndokw, split=False):
     with open(mndokw, 'r') as f:
         mndokeywords = f.read().strip('\n')
     mndokeywords = mndokeywords.split('\n\n')
     charges = np.full(len(mndokeywords), 0)
     mults = np.full(len(mndokeywords), 1)
     for i in range(len(mndokeywords)):
-        status = re.search('kharge=\d+', mndokeywords[i])
+        status = re.search('kharge=-?\+?\d+', mndokeywords[i])
         if status:
             charges[i] = int(status.group().split('=')[1])
         status = re.search('imult=\d+', mndokeywords[i])
@@ -357,10 +375,10 @@ def read_mndokw(mndokw, optmol):
                 mults[i] = 1
             else:
                 mults[i] = imult
-    if optmol is not None:
-        mndokeywords = [mndokeywords[optmol]]
-        charges = [charges[optmol]]
-        mults = [mults[optmol]]
+    if split:
+        for i in range(len(mndokeywords)):
+            with open(f'mndokw_split{i+1}', 'w') as f:
+                f.write(mndokeywords[i])
 
     return mndokeywords, charges, mults
 

{MLatom-2.1.0 → MLatom-2.1.2}/src/MLatom/ML_NEA.py

@@ -93,9 +93,11 @@ def err_print(content: str):  # need to improve
     print('--- ERROR  ERROR  ERROR  ERROR  ERROR  ERROR  ERROR  ERROR  ERROR  ERROR ---')
     print(content)
     print('--- ERROR  ERROR  ERROR  ERROR  ERROR  ERROR  ERROR  ERROR  ERROR  ERROR ---')
+    sys.stdout.flush()
 
 def warning_print(content: str):
     print(f'!!!!! {content} !!!!!')
+    sys.stdout.flush()
 
 # todo: complete the typing of logger decorator
 def logger(log_content: str, logfile: str='out.log') -> Callable[[Any], Any]:
@@ -105,6 +107,7 @@ def logger(log_content: str, logfile: str='out.log') -> Callable[[Any], Any]:
             log_string = '\n\n' + '=' * 90 + '\n'
             log_string += f'{log_content} ( {func.__name__} ) started at {time.ctime()} {time.tzname[0]}'
             print(log_string)
+            sys.stdout.flush()
 
             start_time = time.time()
             result: Any = func(*args, **kwargs)
@@ -114,6 +117,7 @@ def logger(log_content: str, logfile: str='out.log') -> Callable[[Any], Any]:
             log_string = f'{log_content} ( {func.__name__} ) finished at {time.ctime()} {time.tzname[0]} |||| total spent {spend_time:.2f} sec'
             log_string += '\n' + '=' * 90 + '\n\n'
             print(log_string)
+            sys.stdout.flush()
 
             return result
         return wrapped_function
@@ -219,6 +223,7 @@ class LinuxInterface():
         print('*' * 80)
         print(content)
         print('*' * 80)
+        sys.stdout.flush()
     
 
 class XYZ():
@@ -808,6 +813,7 @@ class NewtonXCalc(LinuxInterface):
         self.mv(freq_file, os.path.join(self.NX_path, 'freq.out') )
         nx_inp = '&dat\n nact = 2\n numat = %d\n npoints = %d\n file_geom = geom\n iprog = 4\n' % (atom_num, NEA_point) + \
                  ' file_nmodes = freq.out\n anh_f = 1\n temp = 0.0\n ics_flg = n\n chk_e = 0\n' + \
+                 ' file_out = ini_qv\n file_vib = qvector\n' + \
                  ' nis = 1\n nfs = 2\n kvert = 1\n de = 100\n prog = 6.5\n iseed = 0\n lvprt = 1\n /' 
         with open('initqp_input', 'w') as f:
             f.write(nx_inp)
@@ -821,6 +827,7 @@ class NewtonXCalc(LinuxInterface):
         self.mv(freq_file, os.path.join(self.NX_path, 'freq.out') )
         nx_inp = '&dat\n nact = 2\n numat = %d\n npoints = %d\n file_geom = "geom"\n iprog = 4\n' % (atom_num, NEA_point) + \
                  ' file_nmodes = "freq.out"\n anh_f = 1\n temp = 0.0\n ics_flg = "n"\n chk_e = 0\n' + \
+                 ' file_out = "ini_qv"\n file_vib = "qvector"\n' + \
                  ' nis = 1\n nfs = 2\n kvert = 1\n de = 100\n prog = 6.5\n iseed = 1234\n lvprt = 1\n/\n'
         with open(inp_name, 'w') as f:
             f.write(nx_inp)
@@ -1395,6 +1402,7 @@ class ML_cross_section(CrossSection_calc, ML_calc):
         self._iter_calc_QC(0, self.TD_point)
         rmse = self._train_epoch(self.TD_point)
         print(f'\nRMSE_geom value for {self.TD_point} point: {rmse}')
+        sys.stdout.flush()
         self._train_final(self.TD_point)
         self._collect_ML_result(self.use_model_pred_path)
         self.mix_multiple_file(self.max_state, self.pred_data_path, self.this.gauss.gauss_data_path, self.mix_data_path)
@@ -1411,13 +1419,15 @@ class ML_cross_section(CrossSection_calc, ML_calc):
         # attention: In fact, we have TD_point number: (self.TD_point + 1)
         while current_point < self.TD_point + 1:
             rmse = self._train_epoch(current_point)
-            rRMSE = abs(rmse - prev_rmse) / rmse
+            rRMSE = (rmse - prev_rmse) / rmse
             print(f'ML-NEA iteration {iter_cnt}: train_number = {current_point}; RMSE_geom = {rmse}; rRMSE = {rRMSE}\n')
+            sys.stdout.flush()
             iter_cnt += 1
             # todo 4: print log about the RMSE variation
-            if rRMSE <= threshold:
+            if abs(rRMSE) <= threshold:
                 # calc cross section
                 print(f'ML-NEA iteration ended after {iter_cnt - 1} iteration!')
+                sys.stdout.flush()
                 # todo 5: extract all data and mix data
                 self._train_final(current_point)
                 self._collect_ML_result(self.use_model_pred_path)
@@ -1431,6 +1441,7 @@ class ML_cross_section(CrossSection_calc, ML_calc):
                     print(f'max training number exceeded, use all point for ML-NEA training\n')
                     rmse = self._train_epoch(current_point)
                     print(f'ML-NEA iteration {iter_cnt}: train_number = {current_point}; RMSE_geom = {rmse}; rRMSE = {rRMSE}\n')
+                    sys.stdout.flush()
                     self._train_final(current_point)
                     self._collect_ML_result(self.use_model_pred_path)
                     self.mix_multiple_file(self.max_state, self.pred_data_path, self.this.gauss.gauss_data_path, self.mix_data_path)

{MLatom-2.1.0 → MLatom-2.1.2}/src/MLatom/MLatom.py

@@ -4,10 +4,10 @@
   !---------------------------------------------------------------------------!
   !                                                                           !
   !     MLatom: a Package for Atomistic Simulations with Machine Learning     !
-  !                               Version 2.0                                 !
+  !                               Version 2.1.2                               !
   !                           http://mlatom.com/                              !
   !                                                                           !
-  !                  Copyright (c) 2013-2021 Pavlo O. Dral                    !
+  !                  Copyright (c) 2013-2022 Pavlo O. Dral                    !
   !                           http://dr-dral.com/                             !
   !                                                                           !
   ! All rights reserved. This work is licensed under the                      !
@@ -34,7 +34,7 @@
   ! Pavlo O. Dral, Peikun Zheng, Bao-Xin Xue, Fuchun Ge, Yi-Fan Hou,          !
   ! Max Pinheiro Jr,                                                          !
   ! MLatom: A Package for Atomistic Simulations with Machine Learning         !
-  ! version 2.1, Xiamen University, Xiamen, China, 2013-2021                  !
+  ! version 2.1.2, Xiamen University, Xiamen, China, 2013-2022                !
   !                                                                           !  
   !---------------------------------------------------------------------------!
 
@@ -42,7 +42,7 @@
 
 import os, sys, subprocess, time, shutil, re, copy
 import sliceData, stopper
-import MLtasks, sliceData, stopper
+import MLtasks
 import interface_MLatomF
 import header
 from args_class import ArgsBase
@@ -79,7 +79,15 @@ class MLatomMainCls(object):
         if args.nthreads:
             os.environ["OMP_NUM_THREADS"] = str(args.nthreads)
             os.environ["MKL_NUM_THREADS"] = str(args.nthreads)
-            os.environ["OMP_PROC_BIND"]   = 'true'
+            # os.environ["OMP_PROC_BIND"]   = 'true'
+        try:
+            f_cpu = open('/proc/cpuinfo')
+            for line in f_cpu:
+                if 'AMD'.lower() in line.lower():
+                    os.environ['MKL_DEBUG_CPU_TYPE'] = '5'
+                    break
+            f_cpu.close()
+        except: pass
         
         # Perform requested task
         if args.XYZ2X or args.sample or args.analyze:
@@ -89,8 +97,8 @@ class MLatomMainCls(object):
             sliceData.sliceDataCls(argsSD = args.args2pass)
         elif (args.useMLmodel    or args.createMLmodel or
               args.estAccMLmodel or args.learningCurve or
-              args.AIQM1DFTstar or args.AIQM1DFT or args.AIQM1 or 
-              args.geomopt or args.freq or
+              args.AIQM1DFTstar or args.AIQM1DFT or args.AIQM1 or
+              args.geomopt or args.freq or args.ts or args.irc or
               args.MLNEA):
             MLtasks.MLtasksCls(argsMLtasks = args.args2pass)
         
@@ -109,13 +117,13 @@ class Args(ArgsBase):
             'XYZ2X', 'analyze', 'sample', 'sampleFromSlices', 'mergeSlices','slice',
             'useMLmodel', 'createMLmodel', 'estAccMLmodel', 'learningCurve', 
             'crossSection', # Interfaces
-            'callNXinterface', 'AIQM1DFTstar', 'AIQM1DFT', 'AIQM1', # DEVELOPMENT VERSION
-            'geomopt', 'freq'
+            'AIQM1DFTstar', 'AIQM1DFT', 'AIQM1', # DEVELOPMENT VERSION
+            'geomopt', 'freq', 'ts', 'irc'
         ]
         self.add_default_dict_args(self.task_list, bool)
         self.add_default_dict_args([
             'useMLmodel', 'createMLmodel', 'estAccMLmodel', 'learningCurve','deltaLearn','selfCorrect','MLmodelType',
-            'geomopt', 'freq'
+            'geomopt', 'freq', 'ts', 'irc'
             ],
             bool
         )
@@ -139,6 +147,9 @@ class Args(ArgsBase):
         self.set_keyword_alias('crossSection', ['ML-NEA', 'ML_NEA', 'crossSection', 'cross-section', 'cross_section','MLNEA'])
         self.set_keyword_alias('AIQM1DFTstar', ['AIQM1@DFT*'])
         self.set_keyword_alias('AIQM1DFT', ['AIQM1@DFT'])
+        self.set_keyword_alias('usemlmodel mlmodeltype=ani1ccx', ['ANI-1ccx'])
+        self.set_keyword_alias('usemlmodel mlmodeltype=ani1x', ['ANI-1x'])
+        self.set_keyword_alias('usemlmodel mlmodeltype=ani2x', ['ANI-2x'])
         self.args2pass = []
 
     def parse(self, argsraw):

{MLatom-2.1.0 → MLatom-2.1.2}/src/MLatom/MLatomF_src/A_KRR.f90

@@ -7,7 +7,7 @@
 
 module A_KRR
   use A_KRR_kernel,  only : NtrTot, NtrVal, itrval
-  use dataset,       only : arrayOfArrays, Nprops, NgradXYZmax!, Yest, NitemsTot
+  use dataset,       only : arrayOfArrays, Yprior, Nprops, NgradXYZmax!, Yest, NitemsTot
   use optionsModule, only : option
   use precision,     only : rprec
   use stopper,       only : stopMLatomF
@@ -27,7 +27,7 @@ contains
 !==============================================================================
 subroutine init_KRR()
 !==============================================================================
-  use A_KRR_kernel,  only : getSigma, getC, getD, getNN
+  use A_KRR_kernel,  only : getSigma, getPeriod, getSigmaP, getC, getD, getNN
   use optionsModule, only : option
   use timing,        only : timeTaken
   implicit none
@@ -44,6 +44,11 @@ subroutine init_KRR()
           (trim(option%kernel) == 'exponential')) then
     ! Get user-defined sigma or request for optimizing it
     call getSigma()
+    if (option%periodKernel) call getPeriod
+    if (option%decayKernel)  then
+      call getSigmaP
+      call getPeriod
+    end if
   elseif (trim(option%kernel) == 'Matern') then
     ! Get user-defined c and d or request for optimizing them
     call getNN()
@@ -193,6 +198,7 @@ subroutine optHyper_KRR(hyperOptDataPt, indicesCVoptSplitsPt, doNotCalcStats)
     if (doNotCalcStats) doNotCalcStatsLocal = .true.
   end if
   if (.not. option%CVopt .and. .not. doNotCalcStatsLocal) then
+    if (option%prior/='0') write(6,'(" Prior value = ",F25.13)') Yprior
     ! Predict values using ML model for its own sub-training set
     call calcEst_KRR(Nsubtrain, indicesSubtrain)
     ! Calculate errors of the predicted values
@@ -276,6 +282,7 @@ subroutine train_KRR(doNotCalcStats)
   ! Calculate regression coefficients alpha
   call calcAlpha(NtrTot, Nprops, K(1)%twoDrArr, Ytrain)
   deallocate(Ytrain)
+  if (option%prior/='0' .and. .not. present(doNotCalcStats)) write(6,'(" Prior value = ",F25.13)') Yprior
   
   ! Predict values using ML model for its own training set
   if (present(doNotCalcStats)) then
@@ -396,7 +403,7 @@ subroutine validate1_KRR(NsubtrainPoints, indicesSubtrainPoints, &
     call calcEst_KRR(NvalidatePoints, indicesValidatePoints)
   elseif (present(KvalidateMatrix)) then
     ! Predict values using ML model for its validation set
-    Yest = 0.0_rprec
+    Yest = Yprior
     !$OMP PARALLEL DO PRIVATE(a,i) &
     !$OMP SHARED(Yest,indicesValidatePoints,Kvalidate,alpha,NsubtrainPoints,NvalidatePoints) &
     !$OMP SCHEDULE(STATIC)
@@ -699,6 +706,7 @@ subroutine calcAlpha(NtrainPoints, Nprops, Kmatrix, Yref)
 ! by solving the system of equations (K+lambda*I)*alpha = Yref
 ! Input:  Kmatrix - K, kernel matrix
 !         Yref    - Yref, vector(s) with reference value for various properties
+!                   WARNING: may be modified if prior/=0 is requested
 ! Output: alpha  - alpha, the regression coefficients
 !==============================================================================
   use dataset,       only : Y
@@ -732,6 +740,27 @@ subroutine calcAlpha(NtrainPoints, Nprops, Kmatrix, Yref)
   do jj=1,NtrainPoints
     InvMat(jj,jj) = InvMat(jj,jj) + lambda
   end do
+  
+  ! Get prior (intercept)
+  Yprior = 0.0_rprec  
+  if (option%prior=='mean' .and. Nprops==1) then
+    Yprior = 0.0_rprec
+    do jj=1,NtrTot
+      Yprior = Yprior + Yref(jj,1)
+    end do
+    Yprior = Yprior / NtrTot
+    if (option%debug) write(6,'("   Mean prior = ",F25.13)') Yprior
+  elseif (option%prior/='0') then
+    if (Nprops > 1) call stopMLatomF('Only prior=0 can be used for these calculations')
+    read(option%prior,*) Yprior    
+  end if
+  
+  ! Center the data around prior
+  if (option%prior/='0') then
+    do jj=1,NtrVal
+      Yref(jj,1) = Yref(jj,1) - Yprior
+    end do
+  end if
 
   ! Get vector with regression coefficients alpha:
   ! Solve the system of linear equations (K+lambda*I)*alpha = Yref
@@ -789,10 +818,10 @@ subroutine calcEst_KRR(NestIndices, indicesForEst)
   call initializeAllArrays()
 
   if (option%calcVal) then
-    Yest = 0.0_rprec
+    Yest = Yprior
     !$OMP PARALLEL DO PRIVATE(i) SHARED(Yest,indicesForEst) SCHEDULE(STATIC)
     do i=1, NestIndices
-      Yest(indicesForEst(i)) = Yest_KRR(indicesForEst(i), alpha(:,1))
+      Yest(indicesForEst(i)) = Yprior + Yest_KRR(indicesForEst(i), alpha(:,1))
     end do
     !$OMP END PARALLEL DO
   end if
@@ -918,7 +947,7 @@ recursive subroutine optLambdaLog2Grid(hyperOptDataPt, indicesCVoptSplitsPt)
 
     else
       ! Train, i.e. calculate regression coefficients alpha
-      call getYtrain(Nsubtrain, indicesSubtrain, Nprops)
+      call getYtrain(Nsubtrain, indicesSubtrain, Nprops) ! Necessary to reset values if they are modified by prior
       call calcAlpha(NtrTot, Nprops, K(1)%twoDrArr, Ytrain)
       ! Predict values using ML model
       if (option%onTheFly) then
@@ -948,7 +977,7 @@ recursive subroutine optLambdaLog2Grid(hyperOptDataPt, indicesCVoptSplitsPt)
   lambda = lambda_min
   ! Recalculate regression coefficients alpha
   if (.not. option%CVopt) then
-    call getYtrain(Nsubtrain, indicesSubtrain, Nprops)
+    call getYtrain(Nsubtrain, indicesSubtrain, Nprops) ! Necessary to reset values if they are modified by prior
     call calcAlpha(NtrTot, Nprops, K(1)%twoDrArr, Ytrain)
   end if
 
@@ -1049,7 +1078,7 @@ recursive subroutine optLog2Grid(hyperparameter, hyperOptDataPt, indicesCVoptSpl
             stsize = newsize
           end do
           call cleanUp_KRR()
-          call getYtrain(Nsubtrain, indicesSubtrain, Nprops)
+          call getYtrain(Nsubtrain, indicesSubtrain, Nprops) ! Necessary to reset values if they are modified by prior
           allocate(K(1),stat=Error)
           if(Error/=0)call stopMLatomF('Unable to allocate K')
           allocate(K(1)%twoDrArr(1:NtrTot,1:NtrTot),stat=Error)
@@ -1087,7 +1116,7 @@ recursive subroutine optLog2Grid(hyperparameter, hyperOptDataPt, indicesCVoptSpl
 
     else
       ! Calculate kernel matrix
-      call getYtrain(Nsubtrain, indicesSubtrain, Nprops)
+      call getYtrain(Nsubtrain, indicesSubtrain, Nprops) ! Necessary to reset values if they are modified by prior
       call calcKernel(K(1)%twoDrArr)
       if (.not. option%onTheFly) call calc_Kvalidate(Nsubtrain, indicesSubtrain, &
                                                      Nvalidate, indicesValidate, &

{MLatom-2.1.0 → MLatom-2.1.2}/src/MLatom/MLatomF_src/A_KRR_kernel.f90

@@ -13,6 +13,7 @@ module A_KRR_kernel
   implicit none
   ! Variables
   real(kind=rprec), target :: sigma = 100.0_rprec ! Length-scale
+  real(kind=rprec), target :: sigmap = 100.0_rprec! Length-scale for the periodic part
   integer,          target :: lgSigmaPoints       ! Number of points on a logarithmic grid for search optimal sigma
   real(kind=rprec), target :: lgSigmaL            ! The lowest  value of log2(sigma) to try
   real(kind=rprec), target :: lgSigmaH            ! The highest value of log2(sigma) to try
@@ -25,6 +26,7 @@ module A_KRR_kernel
   real(kind=rprec), target :: lgDL                ! The lowest  value of log2(d) to try
   real(kind=rprec), target :: lgDH                ! The highest value of log2(d) to try
   integer,          target :: nn = 2              ! n in the Matern kernel (nu = n + 1/2)
+  real(kind=rprec), target :: period =  1.0_rprec ! Period
   integer                  :: NtrTot              ! Total number of training points
   integer                  :: NtrVal              ! Number of training points with values
   integer, allocatable     :: itrval(:)           ! Indices of training points with values
@@ -58,6 +60,20 @@ subroutine getSigma()
 
 end subroutine getSigma
 
+subroutine getPeriod()
+  implicit none
+
+  read(option%period,*) period
+
+end subroutine getPeriod
+
+subroutine getSigmaP()
+  implicit none
+
+  read(option%sigmap,*) sigmap
+
+end subroutine getSigmaP
+
 subroutine getC()
   implicit none
 
@@ -464,6 +480,7 @@ function kernelFunction(Xi,Xj)
 end function kernelFunction
 
 function Kfunk(Xsize,Xi,Xj)
+  use constants,     only : pi
   use dataset,       only : distance
   use mathUtils,     only : factorial_rprec
   implicit none
@@ -483,7 +500,13 @@ function Kfunk(Xsize,Xi,Xj)
   elseif (trim(option%kernel) == 'polynomial') then
     Kfunk = (distance(Xsize,Xi,Xj) + c) ** d
   elseif (trim(option%kernel) == 'Gaussian') then
-    Kfunk = exp(-1.0_rprec * distance(Xsize,Xi,Xj) / (2 * sigma ** 2))
+    if (option%decayKernel) then
+      Kfunk = exp(-1.0_rprec * distance(Xsize,Xi,Xj) / (2 * sigma ** 2) - 2.0_rprec * (sin(pi * sqrt(distance(Xsize,Xi,Xj)) / period))**2 / (sigmap ** 2))
+    elseif (option%periodKernel) then
+      Kfunk = exp(-2.0_rprec * (sin(pi * sqrt(distance(Xsize,Xi,Xj)) / period))**2 / (sigma ** 2))
+    else
+      Kfunk = exp(-1.0_rprec * distance(Xsize,Xi,Xj) / (2 * sigma ** 2))
+    end if
   elseif (trim(option%kernel) == 'Laplacian') then
     Kfunk = exp(-1.0_rprec * distance(Xsize,Xi,Xj) / sigma)
   elseif (trim(option%kernel) == 'exponential') then

{MLatom-2.1.0 → MLatom-2.1.2}/src/MLatom/MLatomF_src/MLatomF.f90

@@ -2,10 +2,10 @@
   !---------------------------------------------------------------------------!
   !                                                                           !
   !     MLatom: a Package for Atomistic Simulations with Machine Learning     !
-  !                               Version 2.0                                 !
+  !                             Version 2.1.2                                 !
   !                           http://mlatom.com/                              !
   !                                                                           !
-  !                  Copyright (c) 2013-2021 Pavlo O. Dral                    !
+  !                  Copyright (c) 2013-2022 Pavlo O. Dral                    !
   !                           http://dr-dral.com/                             !
   !                                                                           !
   ! All rights reserved. No part of MLatom may be used, published or          !
@@ -29,7 +29,7 @@
   !                                                                           !
   ! Pavlo O. Dral, Bao-Xin Xue, Fuchun Ge, Yi-Fan Hou,                        !
   ! MLatom: A Package for Atomistic Simulations with Machine Learning         !
-  ! version 2.0, Xiamen University, Xiamen, China, 2013-2021                  !
+  ! version 2.1.2, Xiamen University, Xiamen, China, 2013-2022                !
   !                                                                           !  
   !---------------------------------------------------------------------------!