Skip to content

program.md

Latest commit

 

History

History
56 lines (42 loc) · 2.43 KB

program.md

File metadata and controls

56 lines (42 loc) · 2.43 KB
Raw

PETRA-4 Booster Dynamic Aperture Optimization

Problem Statement

Maximize the dynamic aperture (DA) area of the PETRA-4 booster synchrotron ring by optimizing two sextupole strengths (DBA_k2 and SF_h), subject to chromaticity feasibility constraints (both horizontal and vertical chromaticities must be >= 0).

The optimization uses the mpBAX framework: a Bayesian optimization loop that trains neural network surrogate models of expensive physics simulations, then uses those surrogates to propose new candidate solutions.

What You Can Modify

Only train.py. You can:

  • Tune hyperparameters: learning rates, network sizes, training schedules, number of loops, initial samples, grid resolution, etc.
  • Modify the Algorithm.propose() strategy: replace COBYLA with other optimizers (L-BFGS-B, differential_evolution, dual_annealing), add multi-start optimization, change candidate selection logic
  • Adjust model configuration: model architecture params (dropout, neurons, epochs), training mode, early stopping
  • Implement adaptive strategies: e.g. start coarse then refine, decay learning rate, schedule proposal counts across loops
  • Add exploration mechanisms: propose some points from uncertain regions, use acquisition functions

What You Cannot Modify

  • prepare.py — the physics (lattice, oracles, utilities)
  • evaluate.py — the metric computation
  • program.md, CLAUDE.md — instructions

Metric

DA area in mm^2 of the best feasible solution, evaluated with the real physics oracle. Higher is better. Printed as METRIC: <value> at the end of the run.

Constraint

Total runtime must be under 10 minutes.

Research Directions (by expected impact)

  1. Hyperparameter balance: Finding the right tradeoff between initial samples, number of loops, and model capacity within the time budget
  2. Propose strategy: Replace single-start COBYLA with multi-start optimization, evolutionary methods, or Bayesian acquisition functions
  3. Sampling efficiency: Better initial designs (Sobol, stratified), or adaptive sampling that targets interesting regions
  4. Model improvements: Try different model_type ('fc' vs 'split' vs 'sine'), adjust network depth/width, tune dropout/regularization
  5. Training strategy: retrain vs finetune mode, weight schemes for new vs old data, learning rate schedules
  6. Curriculum learning: Start with coarse grid/few turns, increase fidelity in later loops