Answers are grounded in this paper and cite their sources.
Background Potato yield formation is well understood in terms of light capture, radiation conversion, biomass accumulation and partitioning, but mechanistic crop models often rely on fixed or externally calibrated physiological parameters. This limits their use for population-specific inference, where the practical question is not only which line yielded most, but which physiological route produced that yield. We developed a trainable mechanistic potato dry-matter model implemented as a knowledge-guided neural network (KGNN), where the established crop-growth equations are retained, while small neural submodels infer their physiological parameter values from observations. Methods A multi-source potato panel was assembled with the purpose of constructing and calibrating a KGNN for a Dutch field experiment scenario. The basic modelling unit was a crop sequence or crop time series that represented the longitudinal canopy record for one retained plot, treatment replicate or treatment group. The panel contained 247 sequences: 196 for model fitting and 51 randomly held-out Dutch sequences, stratified by two Dutch environments, for within-population validation. Observed leaf area index (LAI), days after emergence, cumulative photosynthetically active radiation and environment indicators were used to infer effective light capture, radiation-use efficiency (RUE) and a bounded partitioning term. Tuber dry matter was then calculated through the physiological graph. Performance and physiological admissibility were assessed using prediction error, external proxy alignment, graph-consistency checks, structural replays and ±10–20% input perturbations. Results On the 51 held-out sequences, the KGNN achieved a mean absolute error of 1.41 t ha⁻¹ (8.9% of the observed mean), a root mean square error of 1.77 t ha⁻¹ and a relative root mean square error of 11.2%, with near-zero mean bias. An apparent modelled RUE diagnostic aligned strongly with the observed apparent RUE proxy (Spearman r = 0.88). The learned structure followed the imposed production graph, including associations between intercepted radiation and total dry matter (r = 0.94), total dry matter and tuber dry matter (r = 0.99), and total dry matter and the model-conditioned partitioning term (r = −0.54). Input perturbations produced smooth, monotonic responses; prediction changes were larger for cumulative radiation than for LAI. The partitioning latent was the weakest external anchor (r = 0.27) and therefore interpreted cautiously. Consistency in the structural replays and input or latent perturbations provided evidence for physiological-admissibility and knowledge-alignment of the final model despite end-to-end training. Conclusions A mechanistic potato dry-matter graph can be made locally trainable by replacing fixed coefficients with neural parameter-generating functions. The present validation supports physiological inference for unseen records within the target Dutch experimental population, while a new population would require retraining or recalibration.
No takes yet. Share an insight, caveat, or question.
Mhango et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: