SMLR: Surrogate Models for Linear Response

This repository accompanies L. Jin et al., “Surrogate Models for Linear Response”.
It provides user-ready emulators and notebooks for reproducing figures and results.

---

Repository layout

• Dipole_polarizability/ — Emulators for like-particle QRPA (dipole strength & electric dipole polarizability).
• Beta_decay/ — Emulators for charge-exchange QRPA (strengths & $\beta$-decay half-lives).
• figs/ — Jupyter notebooks to recreate paper figures.
• Data (external) — High-fidelity QRPA inputs expected under:
  • dipoles_data_all/ (dipole strengths & $\alpha_D$)
  • beta_decay_data_Ni_80/ (GT strengths & half-lives)

---

Dipole polarizability

Contains emulators for dipole strength and electric dipole polarizability.

Main files

1. dipole_main.py — train EM1 (strengths + polarizability)
2. main_only_alphaD.py — train EM2 (polarizability only)
3. helper.py — shared utilities

Quick start — EM1

# Fast default — trains $n_1 = 10$ model
python main.py

# Customize model and training knobs
python main.py --n 10 --retain 0.5 --fold 2.0     --num-iter 40000 --print-every 500     --n-restarts 5 --seed0 100     --plots save --save-dir runs_em1

More flags: python main.py --help

Outputs (EM1)

• Per-seed: runs_dipole/seed_<SEED>/
  • params_n<N>_retain<R>_seed<SEED>.txt
  • cost_history_seed<SEED>.png / .pdf
  • (optional with --plots save) central_spectrum_fit.png
• Global (top-level):
  • params_best_n<N>_retain<R>.txt — best parameters across all restarts
  • train_set.txt — list of (alpha,beta) grid points used for training

Quick start — EM2

# Fast default — trains $n_2 = 9$ model
python main_only_alphaD.py

# Customize
python main_only_alphaD.py --n 9 --n-restarts 5 --seed0 7     --num-iter 40000 --print-every 500     --plots save --save-dir runs_em2

More flags: python main_only_alphaD.py --help

Outputs (EM2)

• Per-seed: runs_dipole_em2/seed_<SEED>/
  • params_n<N>_seed<SEED>.txt
  • cost_history_seed<SEED>.png / .pdf
  • (optional with --plots save) train_alphaD_compare_iter<K>.png
• Global (top-level):
  • params_<N>_only_alphaD.txt — best parameters across all restarts
  • train_set.txt — list of parameter grid points used for training

---

Beta decay

Contains emulators for charge-exchange QRPA using the beta_decay_data_Ni_80/ dataset.

Main files

1. main.py — train EM1 (strengths + half-lives)
2. main_only_HL.py — train EM2 (half-lives only)
3. helper.py — shared utilities

Quick start — EM1

# Fast default — trains $n_1 = 8$ model
python main.py

# Customize model and training knobs
python main.py --n 12 --retain 0.85 --num-iter 40000 --print-every 500                --n-restarts 3 --seed0 100 --plots save --save-dir runs_em1

More flags: python main.py --help

Outputs (EM1)

• Per-seed: runs_em1/seed_<SEED>/
  • params_n<N>_retain<R>_seed<SEED>.txt
  • cost_history_seed<SEED>.png / .pdf
  • (optional with --plots save) training diagnostics PNGs
• Global (top-level):
  • params_best_n<N>_retain<R>.txt
  • train_set.txt

Quick start — EM2

# Fast default — trains $n_2 = 9$ model 
python main_only_hl.py

# Customize
python main_only_hl.py --n 9 --n-restarts 4 --seed0 42     --num-iter 40000 --print-every 500     --plots save --save-dir runs_em2

More flags: python main_only_HL.py --help

Outputs (EM2)

• Per-seed: runs_em2/seed_<SEED>/
  • params_n<N>_seed<SEED>.txt
  • cost_history_seed<SEED>.png / .pdf
• Global (top-level):
  • params_<N>_only_HL.txt
  • train_set.txt

---

Data

High-fidelity QRPA inputs should be placed at:

• dipoles_data_all/
  • total_strength/strength_<beta>_<alpha>.out
  • total_alphaD/…
• beta_decay_data_Ni_80/
  • strength files and half-life tables used by the Beta-decay emulators

Adjust paths in the scripts if your dataset is stored elsewhere.

---

Figures & notebooks

Notebooks in figs/ reproduce the plots shown in the paper.
They expect trained parameter files produced by the scripts above.

---

EM1 in action (example)

Performance of EM1 on the training set, reproducing Gamow–Teller strength in $^{80}$Ni.

Sweep across $g_0$ (fix $V_0^{\mathrm{is}}$):

Sweep across $V_0^{\mathrm{is}}$ (fix $g_0$):

---

Citation

If you use this code, please cite the associated paper and repository:

L. Jin et al., “Surrogate Models for Linear Response,” 2025.
(Add BibTeX entry here.)