Opal.jl is designed to give data-assimilation and inverse-modelling work one shared route through filters, variational methods and diagnostics. Data assimilation combines a model with observations to estimate what cannot be observed directly, while inverse modelling works backward from observations to infer unknowns. The framework exposes sequential Bayesian filters and variational methods through the same execution, composition and diagnostic pipeline, with advanced capabilities added as composable wrappers. Its stated aim is to address the lack of a unified and extensible framework for these tasks.
The model underneath can also be replaced. In partial-differential-equation workflows, Opal.jl connects to Gridap and GridapROMs, so either a full-order or reduced-order solution can enter the same filtering workflow by replacing the transition model, the part that advances the simulated system. A reduced-order model is a lower-complexity stand-in for a detailed calculation.
The evaluation covered four challenging computational benchmarks spanning chaotic ordinary- and partial-differential-equation dynamics. The reported comparisons came from those four cases.
Different filters, different costs
In the Lorenz benchmark, the ensemble Kalman filter, or EnKF, and the unscented Kalman filter, or UKF, were reported to estimate the true unknowns accurately in both the augmented state and observation spaces. The variational method was also reported as accurate, but it converged more slowly: about four assimilation windows were needed for x4 and the observations to converge, while x1 took longer. Variational simulations were reported to cost much more time than the Bayesian filters.
SIR did not deliver the same result in that benchmark. It was unable to estimate all the unknowns accurately and diverged from the true values after a few time steps. The comparison therefore records different outcomes among methods available through the shared pipeline.
The Van der Pol comparison
In the Van der Pol benchmark, the bias-unaware option failed to recover the true solution. The bias-aware model recovered accurate parameter-and-state estimates and observations, while its innovations—the gaps between model predictions and observations—followed a Gaussian pattern. The result is a within-benchmark comparison between bias-unaware and bias-aware configurations.
The workflow reaches a fluid model
The Navier–Stokes benchmark extended the demonstration to a partial-differential-equation workflow. The EnKF was reported to recover the true parameters, states and observations, with the innovation modelled as zero-mean Gaussian. The reported configuration used 40 warmup iterations followed by 80 data-assimilation time steps.
Memory gains came with an accuracy caveat
The heat-equation case compared a standard reduced-order model with a calibrated version. In the plotted results, the calibrated ROM appeared to greatly outperform the standard ROM in accuracy. The paper also reports that calibration did not uniformly outperform the standard ROM across all heat-equation state variables.
Memory offered a clearer contrast. The ROM surrogate used 0.91 Gb, compared with 13.56 Gb for the finite-element, or FE, model—about 15 times cheaper in memory. The calibrated ROM was approximately 2.7 times cheaper than the FE model in the stated comparison.
The evidence remains a benchmark record
Taken together, the cases support a narrower conclusion than a general ranking of algorithms. Opal.jl provides one route for sequential filters, variational methods, bias-aware configurations and full- or reduced-order model transitions, and the reported examples include several accurate recoveries. But the evidence covers four computational benchmarks. SIR diverged in the Lorenz case, while heat-equation calibration did not improve every state variable. The results should therefore be read as benchmark demonstrations, not as a universal ranking of methods.
The manuscript is labeled arXiv:2608.24265v2 [cs.MS] and dated 26 Aug 2026. Its examples keep filters, variational methods, bias-aware configurations and different model backends inside a composable workflow, while leaving performance beyond the reported cases open.
Paper data and sources
Original title: Opal.jl: a comprehensive, composable framework for data assimilation in Julia
Authors: Nicholas Mueller
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text