System identification stops modeling one plant at a time: matching only the input signal cuts a stirred-tank reactor's prediction error from 1.249 to 0.242

System identification gains a path that does not model each plant separately: one offline pre-trained meta model takes a short input-output record from a new system plus a query input and returns the corresponding output, with no per-system model estimation. The classical pipeline selects a model structure, an identification method and hyper-parameters for each system, repeats those choices from scratch for every new dataset, and discards information from previously identified systems. Across 320 in-distribution simulated Wiener-Hammerstein systems, patching and ensembling reach root-mean-squared errors of 0.048 and 0.038 over a query window of length 1000.

The meta model writes its prediction as a probability distribution, emitting a mean and a standard deviation at every time step, where the standard deviation is that step's predictive uncertainty. A point estimate returns a single output value and says nothing about how trustworthy that value is. Training minimizes the Kullback-Leibler divergence between the predicted and the true conditional distribution, with the distribution parameterized as a multivariate Gaussian with diagonal covariance and the loss approximated by a Monte Carlo average.

The pre-trained meta model can also run in reverse, generating synthetic data that acts as a regularizer for classical identification when real data are scarce. When the real training set is small or weakly informative, a flexible parametric model trained on that data alone tends to overfit and extrapolate poorly. On the coupled electric drives benchmark, whose training set is small, a maximum-likelihood loss weights synthetic samples inversely to their uncertainty, improving the median index and reducing its spread across Monte Carlo runs.

The meta model addresses control and identification researchers who repeatedly identify systems of the same class, in settings where per-system real data are scarce or weakly informative. It has been tested on real data from three public nonlinear identification benchmarks, the Wiener-Hammerstein, cascaded tanks with overflow, and coupled electric drives, where benchmark data serve only as context and never for training. The authors propose scaling to very large meta models with billions of parameters and building distilled architectures for real-time and edge devices.

Meta training requires substantial offline computation on a large synthetic corpus, which classical identification does not need for a single system. The meta model is relatively heavy at inference time and less suitable for embedded or resource-constrained implementations than parsimonious parametric models. When abundant high-quality per-system data are available, classical identification can still reach higher asymptotic accuracy. The meta model returns predictions rather than an immediately interpretable parametric representation, which is a drawback when frequency-response analysis, sensitivity studies, or control design require a compact structured model.