Skip to content

release_notes.md

Latest commit

 

History

History
78 lines (55 loc) · 6.32 KB

release_notes.md

File metadata and controls

78 lines (55 loc) · 6.32 KB
Raw

Release Notes

lsdo_function_spaces 1.1.0 (October 2026)

lsdo_function_spaces 1.1.0 introduces major performance and architectural enhancements to the LSDO lab's function representation and function space framework for Multidisciplinary Design Optimization (MDO). This release features a matrix-free B-spline evaluation engine with exact analytic reverse-mode sensitivities for dynamic parametric coordinates, accelerated and dimension-agnostic surface projections with convex-hull bounding box pruning, automatic memory safety for large-scale queries, and modernized tag-driven CI testing.

Highlights

  • Matrix-Free B-Spline Custom Operations: Replaces sparse COO/BCOO basis matrix assembly with matrix-free local tensor contraction, reducing JAX compilation overhead by 4.6x and accelerating steady-state VJPs by up to 73x.
  • Parametric Coordinates as Design Variables: Full support for parametric coordinates ($u$) as first-class differentiable CSDL variables with exact analytic reverse-mode VJPs, enabling simultaneous optimization of both surface control points and parametric locations.
  • Accelerated Multi-Patch Projections: Prunes 60%–90% of redundant patch evaluations using provable Euclidean distance lower bounds derived from the B-spline convex hull property with zero false negatives.
  • $5.5\times$ Memory Reduction in Grid Search: Replaced 3D broadcasted array allocations with 2D BLAS-3 GEMM matrix multiplications, speeding up grid searches by 4x and reducing peak memory by 82%.
  • Dimension-Agnostic Orthogonal Projection: Replaced 3D-specific cross-products with general orthogonal vector decomposition $|r_{\perp}|^2 = |r|2^2 - \langle r, \hat{d} \rangle^2$, supporting arbitrary dimensions $\mathbb{R}^{d{\text{phys}}}$ ($d_{\text{phys}} \ge 2$).
  • Automatic ~8 GB Memory Safety: Added automatic batch chunking capped at $10^9$ float64 elements (~8 GB RAM) so massive queries run reliably without out-of-memory errors or manual cutoff tuning.

Core Architecture & Features

  • Matrix-Free Tensor Evaluator (b_spline_csdl_custom_ops_v2.py):
    • Evaluates multi-dimensional tensor-product B-spline curves, surfaces, and volumes via localized span contraction (evaluate_b_spline_matrix_free).
    • Computes pointwise reverse-mode VJPs ($\bar{u}_m = \bar{y}_m \cdot J_m$) and scatters $\bar{C}$ locally into active control point spans, eliminating $O(M^2)$ computational graph coupling.
    • Preserves the legacy b_spline_csdl_custom_ops.py module for backward compatibility while wiring b_spline_space.py to use v2.
  • Bounding Box (AABB) Distance Pruning (FunctionSet.project):
    • Precomputes physical bounding boxes $\text{AABB}(C) = [\min C, \max C]$ per patch.
    • Evaluates the $O(1)$ distance lower bound $d(x, \text{AABB}_k) \le \min_u |S_k(u) - x|_2$.
    • Safely skips candidate patches where $d(x, \text{AABB}k) \ge d{\text{best}}(x)$, dramatically speeding up multi-patch geometries like wings and fuselages.
  • Active-Set Newton Stabilization:
    • Two-stage stabilized polish with eigenvalue clamping and Levenberg-Marquardt damping prevents singular matrix updates near boundaries or on degenerate patches.

API Changes & Enhancements

  • Streamlined project Parameter Order:
    • Reordered arguments across Function.project, FunctionSet.project, and Function.refine_projection to place use_line_search immediately after direction, followed by projection_tolerance: 
      def project(self, points, direction=None, use_line_search=False, projection_tolerance=None, ...)
  • Memory Management Cutoffs:
    • Added MAX_GRID_SEARCH_MEMORY_ELEMENTS = 1_000_000_000 (~8 GB RAM).
    • grid_search_evaluation_cutoff and grid_search_subtraction_cutoff default to None and automatically apply the 8 GB limit, while manual cutoffs are retained for backward compatibility.
  • Dimension-Agnostic Refinement:
    • Updated Function.refine_projection to dynamically handle arbitrary physical dimensions (points.shape[-1]) rather than assuming 3D.

Bug Fixes & Stability

  • Direction Normalization: Ensured unit-norm normalization of direction at entry of Function.project, preventing Newton iteration drift during directional ray projections.
  • STEP CAD Entity Parser Fallback: Added robust heuristic scanning in _detect_knot_fields when knot multiplicity and knot lists appear with integer formatting or non-standard entity slots.
  • Rational Surface Detection: Added explicit NotImplementedError when rational NURBS (RATIONAL_B_SPLINE_SURFACE) entities are encountered.
  • CSDL Recorder Initialization: Added _ensure_active_recorder() when loading pickled geometries to prevent unpickling errors when a CSDL recorder is inactive.
  • Python 3.14 / Spawn Multiprocessing Compatibility: Eliminated dependence on module-level globals (global_functions, global_options) in find_best_surface_chunked, passing arguments explicitly via functools.partial to ensure robust multi-worker projection under spawn (Python 3.14, macOS, and Windows).

Documentation & Infrastructure

  • GitHub Actions Tag-Driven CI: Updated .github/workflows/actions.yml to automatically trigger test matrix runs across Python 3.9, 3.10, 3.11, 3.12, 3.13, and 3.14 upon pushing version tags (v*).
  • Sphinx Documentation Clean Build: Verified HTML documentation generation with zero warnings or errors (sphinx-build -q -W).

lsdo_function_spaces 1.0.0 (September 2024)

lsdo_function_spaces 1.0.0 was the first major stable release of the LSDO lab's functional representation framework. This release eliminated Cython build requirements in favor of a 100% pure Python and JAX architecture, integrated with CSDL Alpha for automatic differentiation, expanded unit test coverage, and established modern packaging and documentation.

Highlights

  • Pure Python Architecture: Removed all Cython build and compilation dependencies, enabling seamless pip installation across all platforms.
  • JAX-Accelerated B-Splines: Introduced vectorized B-spline basis evaluation and forward sensitivity calculation in JAX.
  • CSDL Alpha Integration: Native support for CSDL Alpha graph variables, vector operations, and automatic differentiation.
  • Sphinx AutoAPI Documentation: Complete online documentation hosted on ReadTheDocs with API auto-generation.