UNIDE turns equation discovery into cross-PDE transfer: exact term sets survive 80% noise where WSINDy collapses
Synopsis
Recovering a governing equation from noisy space-time trajectories has meant fitting sparse or symbolic regression from scratch on each trajectory, with differentiation amplifying observation noise. UNIDE learns across six one-dimensional PDE families: a blind-spot denoiser that needs no clean targets cleans the trajectory, a shared encoder predicts which candidate terms are active, and weak-form least squares with residual pruning estimates their coefficients. On 900 held-out trajectories at 80% noise it recovers the exact term set for 0.924, against 0.003 for WSINDy.
Interpretation
Equation discovery no longer fits each observed trajectory on its own; one model trained across PDE families decides which terms make up a new trajectory. Sparse and symbolic regression (SINDy, WSINDy, PySR) fit every trajectory from scratch, so experience from one system cannot inform the next and noise robustness must be won one trajectory at a time. Across 900 held-out trajectories and ten noise levels from 1% to 80%, UNIDE has the highest exact term-set match at every noisy level; at 80% noise it reaches 0.924 against 0.003 for WSINDy, while pointwise methods collapse outright.
A neural network decides which terms are active, and a classical least-squares solve decides their coefficients. The denoiser is trained with a blind-spot objective: a random mask hides a few space-time entries and replaces each with a nearby pixel, so the network predicts hidden values from context and needs no clean targets. The encoder maps each frame through a spatial CNN, tags it with time features, processes the sequence with a transformer, and pools with metadata to output per-term activity probabilities; coefficients are then fitted by weak-form least squares on the predicted support, with greedy residual pruning removing terms whose deletion barely changes the fit.
The advantage comes from learned term selection, not from denoising alone. Given the same denoiser, candidate terms, pruning and validation data, weak-form sparse regression still has lower exact match in every setting. At 10% noise UNIDE reaches 0.960 exact match against 0.787 for the matched sparse regression; at 80% noise it is 0.924 against 0.834; under coefficient extrapolation it is 0.906 against 0.642.
Perspective
The conclusions rest on synthetic, one-dimensional, periodic and fully observed space-time fields; the term structures of the six PDE families all appear in training, and each noise level uses its own trained specialist. The authors propose testing held-out operator combinations and PDE families, physical measurements, and higher-dimensional domains as next steps.
On Burgers the encoder substitutes one transport term for another, which removal-only pruning cannot repair, leaving that family's exact match at 0.23 at 1% noise and 0.00 at 70%; on KdV the small fixed term added for numerical stability can be pruned away, on up to 29 of 150 trajectories. With weights and thresholds frozen at the 10% setting, exact match at 80% noise falls to 0.489, and coefficient accuracy trails WSINDy up to 5% noise.
