RESEARCH QUESTION: Should JuliaFEM use hand-calculated derivatives or AD?
Created comprehensive benchmark comparing:
- Manual: Hand-calculated derivatives (traditional FEM)
- AD: Tensors.jl gradient() (automatic differentiation)
RESULTS (AMD Ryzen 9, Julia 1.12.1):
- Manual: 8.7 ns, 0 allocations
- AD: 268.1 ns, 0 allocations
- AD is 30× SLOWER than manual
KEY FINDINGS:
✅ Both achieve zero allocations (Tensors.jl is well-optimized)
❌ AD has 30× compute overhead from dual number arithmetic
⚠️ In assembly loops: millions of calls = 10+ seconds extra per solve
RECOMMENDATION:
- Keep manual derivatives for common elements (Tet10, Hex8, Quad4, etc.)
- Use AD for prototyping and rare elements
- Unit test manual vs AD to catch errors
- Future: Generate derivatives symbolically (Symbolics.jl)
WHY NOT AD EVERYWHERE?
Assembly is hottest path in FEM. 30× overhead = unacceptable for
production code. Users will notice the performance difference.
WHY NOT ABANDON AD?
- Excellent for prototyping
- Required for exotic bases (NURBS)
- Perfect for unit testing manual derivatives
- Zero allocations impressive
Files:
- benchmarks/tet10_derivatives_benchmark.jl (runnable benchmark)
- docs/benchmarks/shape_function_derivatives_ad_vs_manual.md (analysis)
Dependencies added: BenchmarkTools
This answers the research question definitively with data.
New testing strategy: Educational tests using Literate.jl
Core principles:
- Tests are primary teaching material (not just validation)
- Literate.jl generates docs from test files (always synchronized)
- Structured progression: fundamentals → linear → nonlinear → advanced
- Fast tests (< 5 min unit, < 30 min full suite)
- Target: 99% code coverage
Test hierarchy:
- tutorials/ - Literate.jl files (test + documentation)
- unit/ - Fast isolated function tests
- verification/ - Known analytical solutions
8-week implementation roadmap:
Week 1: Infrastructure (Literate.jl setup)
Week 2-3: Core tutorials (10-15 fundamental topics)
Week 4-5: Advanced tutorials (contact, mortar)
Week 6: Unit tests (fill coverage gaps → 99%)
Week 7: Verification tests (validate correctness)
Week 8: Polish and publish documentation
Philosophy: 'Tests are not a chore - they teach users how to use JuliaFEM.'
Ready to start Phase 1 implementation.
Document the 49 failing tests with clear categorization:
- 14 tests need HDF5 (aster_read_mesh)
- 30 tests have API signature mismatches
- 2 tests already fixed (Analysis export, Statistics)
Includes 4-phase action plan with time estimates.
Good news: Core architecture is sound (package loads, 5 tests pass).
Failures are mechanical API compatibility issues from Julia evolution
(0.6 → 1.12 over 6 years), not fundamental problems.
- Add Tensors and Calculus to Project.toml dependencies
- Add basis includes to src/JuliaFEM.jl (Phase 1 integration)
- Fix FEMBasis. namespace references → use JuliaFEM namespace
- Update create_basis.jl: AbstractBasis (not FEMBasis.AbstractBasis)
Status: Basis files load, but conflict with FEMBase expectations
Next: Need to consolidate FEMBase or work around AbstractElement type constraints
This is expected during consolidation - we're bridging two systems.
It looks document generation proceduce has slightly changed.
docs/Project.toml is defining dependencies for document generation and
they are not explicitly given in `travis.yml`.
Let's use Literate.jl to automatically generate usage examples.
* Automatically generate documentation from other packages (first try to include each package's docs/src/index.md, but if that fails, then use README.md to introduce the package).
* Add example how to calculate local element matrices.
* Add example how to perform 2d contact analysis.
matplotlib cannot be installed during the generation of documentation,
ssl error. Use static images in documentation instead of automatically
generated ones.
A lot of old files from old documentation systems etc. is in package.
These are now removed or moved. Old notebooks are in docs/tutorials.
This PR closes issue #124.
Source code related to read and parse ABAQUS .inp files is now living in
it's own repository `AbaqusReader.jl` and in this commit we cleanup the
same files from this repository.
- add AbaqusReader to .travis.yml because it's not registered package yet
- initialize Mesh from AbaqusReader.jl dict
- remove ABAQUS tests and files moved to AbaqusReader.jl
- remove references to old module Abaqus
- move ABAQUS code to preprocess.jl (what is left)
- close issue #122
- close issue #55
* running v0.6 conversion code proposed by @ovainola in #108.
* change travis so that build is done using 0.6
* documentation is build from 0.6
* fix most of deprecation warnings
* fix test to pass 0.6
- remove some automatically generated stuff not should even be in
repository
- set up lint + Documents.jl in same way it is defined in freshly started projects
- add lint + doctest to after_success so that build pass, these needs to be fixed later
- build is failing on nightly (0.7) but it's ok for release (0.5.2)
- Documenter.jl supports doctests, so this closes least #23
- build system is now on Travis-CI completely, so this closes also #68
force vector, autodiff takes care of linearization
- elasticity equations are now solved using e.g. principle of minimum
potential energy. syntax is quite good, see notebook.
- updated how to interpolate fields, by introducing function spaces.
syntax is now good. still have to figure out how to do time derivatives
- etc. etc. tutorial is broken at the moment, i took of get_lhs and
get_rhs because they didn't really work.
Jukka Aho