v3.1.2

The highlights of this release are a new SR-based data assimilation module, a major overhaul of the data assimilation framework with support for both iceflow backends, and new optimization solvers (MUON, SOAP, improved CG-Newton and Trust Region).

Improvements and Updates

Modules

• data_assimilation:

  • Compatibility: new dispatch layer for use with both emulated and unified iceflow
  • Fix: early stop with patience for the optimization loop

• data_assimilation_SR (new):

  • SR-based inversion with flexible misfit terms (Huber, L²) and regularization (L2, Laplacian-squared)
  • Normalization: cost terms normalized by domain area (ice mask)
  • Fix: boundary effects in mask-based regularization

• iceflow:

  • New: no_inflow boundary condition
  • New: NiceNet and CNO emulator architectures
  • Schema v3 for pretrained model artifacts with improved loading and validation
  • Improved double-precision support
  • Sliding: new u_ref / tau_ref convention for Weertman (more interpretable); Budd and Coulomb slidingco definitions aligned with Weertman; new use_mask_gr flag to force zero friction on floating areas

• pretraining:

  • Overhaul: warmup phase, cleaner normalization layer, model save/load
  • Checkpoint resumption

• optimizers:

  • New: MUON and SOAP optimizers
  • Improved CG-Newton and Trust Region solvers
  • L-BFGS-B: improved bound constraints and memory updates

• live_dashboard (new):

  • Real-time monitoring during runs

Code maintenance

• Optuna is now an optional dependency (pip install igm-model[optuna])
• Early stop with patience for training loops
• Profiler option added

---

Contributors: @shrrosier, @brfi3983, @tgregov, @jouvetg, @gillian-smith, @tancredeleger.