Parameter Optimization

run_calculation(::Val{:optimization}, comm, conf::Config; rank=0, nranks=1)

Performs the optimization using gradient descent of an effective Hamiltonian model using the specified configuration. The routine iteratively adjusts model parameters to minimize the loss over a training set, while optionally validating performance on a separate validation set.

Workflow

1. Configuration Sampling

   • Partition configuration indices into training (train_config_inds) and validation (val_config_inds) sets.
   • Distribute indices among parallel ranks.

2. Translation Vectors (if hr_fit)

   • Retrieve translation vectors Rs for high-rank fitting, used in constructing the Hamiltonian basis.

3. Training Data Preparation

   • Load the training structures (train_strcs) from the selected configurations.
   • Construct the corresponding basis functions (train_bases).
   • Initialize an EffectiveHamiltonian (ham_train) with the training data.

4. Validation Data Preparation

   • Load validation structures (val_strcs) and construct basis functions (val_bases).
   • Initialize an EffectiveHamiltonian (ham_val) for validation, including only the relevant data points.

5. Data Loader and Optimizer Initialization

   • Create a DataLoader (dl) containing training and validation datasets.
   • Initialize a gradient descent optimizer (optim) using the extracted training parameters.

6. Profiler Setup

   • Create a HamsterProfiler (prof) to record losses, timings, and parameter evolution during optimization.

7. Model Optimization

   • Execute optimize_model!, iterating over training and validation data to update model parameters.
   • On rank 0, optionally write optimized parameters to output files.

Required Inputs

• Structural configurations:
  • train_mode = md: see get_xdatcar and get_sc_poscar
  • train_mode = pc: see get_poscar
• Eigenvalue training data: see get_train_data
• Eigenvalue validation data (optional): see get_val_data If not provided, the training data is split according to val_ratio.

Input Tags (Optimizer)

• Configuration tags that specify how optimization is performed (see here).

Output Files

• hamster.out — standard Hamster output.
• hamster_out.h5 — HDF5 file containing:
  • Loss values over iterations
  • Timings