dfsc Ecosystem Maturity
dfsc is developed as a specialized tool-algorithm library for differentiable
Mittag-Leffler spectral learning. Its design borrows the problem/algorithm
separation common in mature scientific-computing ecosystems, while keeping a
strict scope: known or user-supplied spectral propagators with differentiable
fractional parameters.
Current Maturity Layer
Passing the repository audit means the research artifact is internally consistent; it does not mean that dfsc has reached the release maturity or equation coverage of long-lived public solver ecosystems.
| Layer | Current state | Remaining gap |
|---|---|---|
| Package identity | Single public namespace: dfsc |
Public release and hosted docs |
| Internal release gate | 20/20 pre-release checks pass | External adoption and public history |
| Core primitive | Batched PyTorch real-negative MLSL plus controlled moderate-radius complex evaluation | Large complex sectors and certified complex error control |
| Algorithms | Direct, stable, forced, Lanczos, controlled Arnoldi, FFT Caputo-L1 history, operator, graph, generalized operator, semilinear Picard, and L1 fallback | Contour/rational methods and online fast-memory algorithms |
| Problem interface | Spectral, graph, assembled operator, forced, semilinear, linear Caputo L1, and four scoped application templates | Standardized inverse and variable-order problem templates |
| Solution object | Values, time axis, retcode, warnings, stats, and selection diagnostics | Serialization and richer error estimates |
| Numerical routing | AutoDFSC and diagnostic Mittag-Leffler evaluation |
Calibrated rules over wider verified regimes |
| Autograd/GPU | Tested for alpha/beta, matrix-free, nonlinear, and FFT-history gradients on CUDA | Larger multi-GPU and mixed-precision profiling |
| Reliability contract | Validated-domain, convergence, gradient, and empirical error metadata | Rigorous global error bounds |
| Data evidence | One experimental H-actin SPT condition with provenance, checksum, five trajectory splits, and model comparisons | Explicit redistribution permission and independent physical modalities |
| Experimental orders | Variable-order and distributed-order wrappers | Systematic theory and large-scale validation |
Intended Differentiation
dfsc is not yet a general fractional solver library. Its primary path remains
Mittag-Leffler spectral propagation, while the constant-order linear Caputo L1
solver establishes the first explicit history-aware fallback path.
The main differentiator is not solver breadth. It is the ability to expose that propagator as a differentiable layer, algorithm object, and solution-producing workflow inside PyTorch.
The current center of application coverage is anomalous transport and linear fractional relaxation. Around that center, the same primitive now supports assembled finite-element systems, graph diffusion, and controlled non-self-adjoint advection--diffusion. Each template carries machine-readable assumptions and limitations rather than implying general domain coverage.
Next Maturity Milestones
- Resolve redistribution permission for the integrated SPT data and add a second benchmark with an explicit open-data license.
- Add an ecosystem comparison script against at least one PINN/fPINN workflow and one fractional numerical solver on the same task.
- Add block Krylov/Arnoldi and preconditioning strategies for larger batches and stiff spectra.
- Add contour/rational complex actions and an online fast-memory CQ/SOE time stepper before promoting broader solver claims.
- Promote experimental order wrappers only after accuracy and stability tests.
- Publish a versioned source archive, wheel, and hosted API documentation.