TerraKAN: Fractional Kolmogorov-Arnold Transformer with Padé Mittag-Leffler Basis and Biokinetic Mixture-of-Experts

In this work, I introduce TerraKAN (Frac-ML-KAT-MoE), a neuro-symbolic Scientific Machine Learning (SciML) architecture that I designed, mathematically verified, and calibrated to solve the ill-posed inverse problem of soil bioindication and long-horizon self-purification forecasting from sparse (6–8 points), irregular (up to 50% missingness), and noisy (25–40% multiplicative error) enzymatic and respiratory field measurements without direct chemical assays of the contaminant. Whereas polynomial Kolmogorov-Arnold Networks (ChebPIKAN) suffer from catastrophic Runge-type divergence during blind extrapolation on t ∈ [150,365] days(R2 =−8.13) and classical Bayesian ODE solvers (MAP-ODE regularized toward a Soil Atlas prior) fail to reconstruct unmeasured contaminant loads (R2 = −0.191) due to ignoring non-Markovian micropore trapping (α < 1), I unify four analytical primitives into an end-to-end differentiable model: (1) a causal Fractional Kolmogorov-Arnold Transformer encoder (Frac-KAT) with a singular Abel–Riemann–Liouville memory kernel K(t,s) ∝ (t −s)α−1/Γ(α) and strict weight tying of the memory exponent α; (2) a pole-free global [3/4] Pad´e Mittag-Leffler spectral basis (MittagLefflerPade34) matching exact Maclaurin (z → 0) and Poincar´e (z → ∞) asymptotics with denominator Q4(z) ≥ 1.001; (3) a log-exponential MultKAN symbolic discovery node pairedwith a sparse Top-2 biokinetic Mixture-of-Experts (PhysicsExpertBank) separating Haldane substrate inhibition from Monod nutrient limitation; and (4) a dynamic memory-aware soil health index IBSPα(t). Across N = 500 held-out test scenarios on t ∈[150,365] days, TerraKAN achieves R2 = 0.908±0.019 (RMSE = 0.181EC50) on the unmeasured contaminant load ˜P(t) and R2 =0.886±0.016 (RMSE=6.88%) on IBSPα(t) at 0.50 ms per field (10.1× faster than MAP-ODE). Finally, I validate the frozen checkpoint (terrakan_v4_frozen.pt) on a 36-vessel laboratory microcosm protocol across sandy, loamy, and peat soil series, achieving 100% separation of healthy controls, nutrient-poor clean soils, and contaminated regimes without control-flag leakage.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-12
DOI
https://doi.org/10.5281/zenodo.24168982
Primary Topic
Model Reduction and Neural Networks
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

TerraKAN: Fractional Kolmogorov-Arnold Transformer with Padé Mittag-Leffler Basis and Biokinetic Mixture-of-Experts

Ilya Kakhno
Zenodo (CERN European Organization for Nuclear Research)
Model Reduction and Neural Networks
preprint

TerraKAN: Fractional Kolmogorov-Arnold Transformer with Padé Mittag-Leffler Basis and Biokinetic Mixture-of-Experts

Ilya Kakhno
preprint en

Abstract

In this work, I introduce TerraKAN (Frac-ML-KAT-MoE), a neuro-symbolic Scientific Machine Learning (SciML) architecture that I designed, mathematically verified, and calibrated to solve the ill-posed inverse problem of soil bioindication and long-horizon self-purification forecasting from sparse (6–8 points), irregular (up to 50% missingness), and noisy (25–40% multiplicative error) enzymatic and respiratory field measurements without direct chemical assays of the contaminant. Whereas polynomial Kolmogorov-Arnold Networks (ChebPIKAN) suffer from catastrophic Runge-type divergence during blind extrapolation on t ∈ [150,365] days(R2 =−8.13) and classical Bayesian ODE solvers (MAP-ODE regularized toward a Soil Atlas prior) fail to reconstruct unmeasured contaminant loads (R2 = −0.191) due to ignoring non-Markovian micropore trapping (α < 1), I unify four analytical primitives into an end-to-end differentiable model: (1) a causal Fractional Kolmogorov-Arnold Transformer encoder (Frac-KAT) with a singular Abel–Riemann–Liouville memory kernel K(t,s) ∝ (t −s)α−1/Γ(α) and strict weight tying of the memory exponent α; (2) a pole-free global [3/4] Pad´e Mittag-Leffler spectral basis (MittagLefflerPade34) matching exact Maclaurin (z → 0) and Poincar´e (z → ∞) asymptotics with denominator Q4(z) ≥ 1.001; (3) a log-exponential MultKAN symbolic discovery node pairedwith a sparse Top-2 biokinetic Mixture-of-Experts (PhysicsExpertBank) separating Haldane substrate inhibition from Monod nutrient limitation; and (4) a dynamic memory-aware soil health index IBSPα(t). Across N = 500 held-out test scenarios on t ∈[150,365] days, TerraKAN achieves R2 = 0.908±0.019 (RMSE = 0.181EC50) on the unmeasured contaminant load ˜P(t) and R2 =0.886±0.016 (RMSE=6.88%) on IBSPα(t) at 0.50 ms per field (10.1× faster than MAP-ODE). Finally, I validate the frozen checkpoint (terrakan_v4_frozen.pt) on a 36-vessel laboratory microcosm protocol across sandy, loamy, and peat soil series, achieving 100% separation of healthy controls, nutrient-poor clean soils, and contaminated regimes without control-flag leakage.

Zenodo (CERN European Organization for Nuclear Research)
Belarusian State University (BY)
Model Reduction and Neural Networks
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon