warn for too small energies in train/test.xyz

doc/nep/input_files/train_test_xyz.rst

@@ -66,5 +66,6 @@ Tips
 * The minimal number of atoms in a configuration is 1.
   The user is responsible for choosing a sensible reference energy when preparing the energy data.
   But this is not crucial as the absolute energies are not relevant in the present context.
+  However, because NEP training uses single precision, accuracy will be lost if any reference energy is smaller than -100 eV/atom. The code will give a warning message in this case.
 * The energy and virial data refer to the total energy and virial for the system.
   They are not per-atom but per-cell quantities.

src/main_nep/fitness.cu

@@ -33,10 +33,6 @@ Get the fitness
 
 Fitness::Fitness(Parameters& para)
 {
-  print_line_1();
-  printf("Started reading train.xyz.\n");
-  print_line_2();
-
   int deviceCount;
   CHECK(cudaGetDeviceCount(&deviceCount));
 
@@ -72,10 +68,6 @@ Fitness::Fitness(Parameters& para)
   std::vector<Structure> structures_test;
   has_test_set = read_structures(false, para, structures_test);
   if (has_test_set) {
-    print_line_1();
-    printf("Started reading test.xyz.\n");
-    print_line_2();
-
     test_set.resize(deviceCount);
     for (int device_id = 0; device_id < deviceCount; ++device_id) {
       print_line_1();

src/main_nep/structure.cu

@@ -363,6 +363,20 @@ read_exyz(const Parameters& para, std::ifstream& input, std::vector<Structure>&
     ++Nc;
   }
   printf("Number of configurations = %d.\n", Nc);
+
+  for (const auto& s : structures) {
+    if (s.energy < -100.0f) {
+      std::cout << "Warning: \n";
+      std::cout << "    There is energy < -100 eV/atom in the data set.\n";
+      std::cout << "    Because we use single precision in NEP training\n";
+      std::cout << "    it means that the reference and calculated energies\n";
+      std::cout << "    might only be accurate up to 1 meV/atom\n";
+      std::cout << "    which can effectively introduce noises.\n";
+      std::cout << "    We suggest you preprocess (using double precision)\n";
+      std::cout << "    your data to make the energies closer to 0." << std::endl;
+      break;
+    }
+  }
 }
 
 static void find_permuted_indices(std::vector<int>& permuted_indices)
@@ -474,6 +488,9 @@ bool read_structures(bool is_train, Parameters& para, std::vector<Structure>& st
       has_test_set = false;
     }
   } else {
+    print_line_1();
+    is_train ? printf("Started reading train.xyz.\n") : printf("Started reading test.xyz.\n");
+    print_line_2();
     read_exyz(para, input, structures);
     input.close();
   }