These figures come from the SGA-PDE worked example.
Each image and its reading note is taken directly from that run's
full HTML report. Field predictions come from forward
integration; derivative predictions are evaluated on the search target.
In a notebook, model.report("out/report") displays the complete report inline.
The same returned object lists the saved figures and any rendering notes.
Convergence. Read the best native score reached at each search iteration.Parity. Compare predicted and observed targets against the diagonal of exact agreement.Equation. Read the published equation with its fitted coefficients and active terms.Equation Tree. Read the operators and derivatives that compose the published equation.Residual. Read the distribution and spatial pattern of predicted minus observed target values.Coefficient Bar. Compare the magnitudes and signs of the active fitted coefficients.Field Comparison. Compare the measured field with forward integration of the discovered equation and their difference.Pde Residual Field. Compare the fitted derivative target with the equation prediction and their residual, without forward integration.Time Slices. Compare measured and integrated fields at the labeled observation times.Error Heatmap. Locate the signed field error from forward integration at the labeled physical coordinates.Offspring Diversity. Distinct candidate-expression count among this generation's evaluated offspring (the staged batch, post-dedup; failed evaluations count too). A collapsing count signals premature convergence / loss of search diversity. NOT a population statistic — the count can exceed the population size.Complexity Evolution. Mean number of STRidge-selected (non-zero) terms across this generation's valid offspring — a sparsity / bloat monitor for the evolving PDE candidates. All-invalid generations appear as gaps (nothing was measured).Fitness Spread. Per-generation MEAN AIC over the surviving population's finite scores (lower is better). The platform 'convergence' plot draws the single best AIC; this curve shows the level of the whole surviving population — under elitist truncation it declines alongside the best curve, and the distance between the two is how far the population trails its optimum. Generations with no finite-scored survivors (e.g. an all-invalid start) appear as gaps (their +inf sentinel is masked).Surrogate Training Curve. FieldModel surrogate training loss vs epoch (log-y): train and — when a validation split exists — validation MSE, with the best-validation epoch marked. The autograd derivatives (u_x, u_t) the GA reads from come from this network. Absent (No data panel) for finite-diff mode or when the surrogate was pre-trained (user-supplied field_model), not trained by the platform builder.Genome Tree. Raw GP genome of the best individual — operator names (*, ^2, d) as the genetic operators produced them, pre-canonicalization. Shows all evolved terms before STRidge pruning, so its term count may exceed the discovered equation's (Expression Tree). Unavailable for results loaded from disk (no live population).
