DyRC: Dynamics-Informed Reservoir Computing

• DyRC is a reservoir computing method that integrates dynamical systems principles by designing reservoirs whose internal evolution mirrors target or physical dynamics.
• It employs data-derived topologies, heterogeneous timescales, and generalized synchronization to improve prediction accuracy, attractor reconstruction, and bifurcation analysis.
• Physical and abstract implementations, such as nanowire networks and visibility-graph constructions, demonstrate DyRC's ability to harness intrinsic substrate dynamics for efficient nonlinear processing.

Dynamics-Informed Reservoir Computing (DyRC) denotes a family of reservoir-computing formulations in which the reservoir, the readout, the training scheme, or the physical substrate is designed and interpreted through dynamical-systems structure rather than treated as a purely generic random recurrent network. In standard reservoir computing, the reservoir update is typically written as $$\mathbf{x}(t+\Delta t)=f\!\left(W\mathbf{x}(t)+W_{\rm in}\mathbf{u}(t)\right)$$, with fixed recurrent weights $W$ and nonlinearities injected through the node activation $f$. DyRC departs from this picture by treating the reservoir as a dynamical representation whose internal evolution may be governed by synchronization manifolds, heterogeneous timescales, data-derived graph structure, or explicit physical laws. In physically grounded DyRC, the analogous evolution can instead be written abstractly as $$\mathbf{x}(t+\Delta t)=M(t)\,W_{\rm in}\mathbf{u}(t)$$, where $M(t)=M\!\big(t;\mathbf{x}(t);\mathbf{u}(t)\big)$ is a compact representation of full physical evolution rather than a conventional static weight matrix (Xu et al., 22 May 2025). Across the literature, this viewpoint appears both explicitly, as in dynamic physical reservoirs and visibility-graph reservoir construction, and implicitly, as in generalized-synchronization theory, multiscale leak-rate design, parameter-aware autonomous modeling, and physical substrates whose own nonlinear fading-memory dynamics perform the computation (Geier et al., 25 Jul 2025, Ookubo et al., 2024, Hart, 2021).

1. Conceptual scope and distinction from conventional reservoir computing

DyRC is unified less by a single architecture than by a common modeling stance: the reservoir is treated as a dynamical medium whose structure should reflect either the target dynamics or the substrate dynamics. In one explicit formulation, physical neuromorphic nanowire networks are described as dynamic reservoirs in which nanowires are nodes, memristive junctions are edges, and the reservoir state emerges from Kirchhoff-constrained voltage redistribution together with time-evolving conductances. The reservoir is therefore not a static recurrent matrix plus pointwise nonlinearity, but a self-organized circuit whose node activities and edge dynamics are coupled through nonlinear nano-electronic circuit elements (Xu et al., 22 May 2025). In another explicit formulation, DyRC with visibility graphs replaces the usual randomly generated reservoir graph by a graph inferred directly from the training time series via the visibility-graph criterion, so that the reservoir topology is directly informed by the specific dynamics of the prediction task under study (Geier et al., 25 Jul 2025).

This broader scope means that DyRC includes both structural and functional departures from standard RC. Some DyRC methods alter the internal reservoir architecture, for example by introducing diverse intrinsic timescales or by constructing the recurrent graph from the data itself. Others leave the reservoir dynamics conventional but make the readout dynamics-informed, as in generalized readout schemes derived from generalized synchronization. Still others use an existing physical process—such as traffic flow, memristive devices, hydrodynamically coupled oscillators, or coupled microelectromechanical resonators—as the reservoir substrate, with learning confined to the output layer (Ando et al., 2019, Armendarez et al., 2023, Heuthe et al., 9 Jan 2026, Farah et al., 6 Jan 2026). A plausible implication is that DyRC is best understood as an organizing category for reservoir methods that seek structure-function alignment, rather than as a single algorithmic recipe.

2. Dynamical-systems foundations: synchronization, embedding, and invariant geometry

A central theoretical basis for DyRC is generalized synchronization. In the deterministic setting, a reservoir computer with the Echo State Property admits a generalized synchronization between the input dynamical system and the dynamics in reservoir space, and under suitable smoothness and contraction assumptions this synchronization can be $C^1$. In the special case of a linear reservoir map with the Echo State Property, the generalized synchronization is generically an embedding, a result that the thesis states admits Takens’ embedding theorem as a special case (Hart, 2021). This establishes a rigorous foundation for treating the reservoir state as a coordinate representation of the driving dynamics rather than as an opaque feature vector.

The geometric content of this viewpoint is developed further in work on model attractors in high-dimensional neural-network reservoirs. There the autonomous reservoir is treated as a high-dimensional dynamical system in which the true attractor is reconstructed on a low-dimensional invariant set inside reservoir space. Using the Hénon map as the test system, the study shows that the leading Lyapunov exponent of the actual system is recovered throughout the tested spectral-radius range, and that after restricting the analysis to the tangent space of the inferred inertial manifold, the negative Lyapunov exponent is recovered as well. The paper’s formulation emphasizes that the reservoir’s invariant set need not be an attractor in the full ambient space; rather, the true dynamics may live on an embedded attractor with transverse directions that are strongly contracting or weakly unstable (Kobayashi et al., 16 Sep 2025). This directly supports a DyRC interpretation in which faithful modeling is judged not only by short-term prediction but by whether the reservoir preserves tangent dynamics and invariant geometry.

This dynamical-systems framing also motivates a reinterpretation of the output map. If the driven reservoir satisfies $\mathbf r_t=\mathbf f(\mathbf x_t)$ on the attractor, then future prediction is governed by a generally nonlinear map $$\mathbf h(\mathbf r_t)=\boldsymbol\phi^\tau(\mathbf f^{-1}(\mathbf r_t))$$. Standard linear readout, $\hat{\mathbf y}_t=W\mathbf r_t$, then becomes only the first term of a Taylor approximation to the true state-to-output relation. Reservoir computing with generalized readout therefore augments the readout by nonlinear combinations of reservoir variables, for example the quadratic form $$\hat{\mathbf y}_t = W\mathbf r_t + \mathbf r_t^T W^{\mathcal Q}\mathbf r_t$$, while remaining within a linear learning framework because the model is still linear in the trainable parameters (Ookubo et al., 2024). In DyRC terms, the readout is informed by the expected smooth functional relation between target and reservoir dynamics.

3. Dynamics-informed design principles

One major DyRC design principle is to align the reservoir’s internal timescale structure with the target system’s timescale structure. The DTS-ESN model implements this by replacing identical leaky-integrator neurons with heterogeneous leaky-integrator neurons whose leak rates satisfy $W$0. In the reported experiments on four chaotic fast-slow systems, the best DTS-ESN performance is better than the best LI-ESN performance for all four systems in the one-step-ahead task, and the DTS-ESN achieves the largest valid time in all four systems under the tested settings in the long-term closed-loop task (Tanaka et al., 2021). The authors further show that neurons with large $W$1 contribute mainly to prediction of fast subsystem variables, whereas neurons with small $W$2 contribute mainly to prediction of slow subsystem variables. This gives DyRC a clear operational principle: if the target dynamics are multiscale, distribute reservoir memory across multiple internal timescales.

A second design principle is to infer reservoir topology from the signal itself. DyRC with visibility graphs constructs the reservoir adjacency matrix from a training segment of the time series by connecting two samples when they satisfy the visibility criterion $W$3 for every intermediate point. The resulting graph is then normalized to spectral radius $W$4 and used directly as the reservoir matrix. In the Duffing-oscillator experiments, the variant DyRC-VG 16, which uses every 16th data point for graph construction, has the lowest mean error and the smallest variance across repeated implementations, and it outperforms an Erdős–Rényi graph of the same size, spectral radius, and comparable density (Geier et al., 25 Jul 2025). This suggests that task-informed topology can reduce both prediction error and realization-to-realization variability.

A third principle is to enrich the readout rather than only the reservoir. In Lorenz-chaos experiments, quadratic-form RC significantly outperforms a standard linear ESN even with very small reservoirs, including $W$5, and for $W$6 the reported RMSE can be about $W$7 times that of the linear ESN. In closed-loop autonomous mode, the presented QRC cases predict Lorenz dynamics for about 8 Lyapunov times, compared with roughly 3 Lyapunov times for the linear ESN, while also showing improved robustness to random reservoir realization (Ookubo et al., 2024). The key DyRC point is not simply that nonlinear readouts are useful, but that their form is derived from generalized-synchronization theory rather than added ad hoc.

4. Physical DyRC: computation by intrinsic substrate dynamics

Physical implementations occupy a central place in DyRC because they instantiate the claim that the dynamics themselves are the computational substrate. In physical neuromorphic nanowire networks, the reservoir is a graph of nanowires and memristive junctions whose conductance-like internal “weights” evolve in time. Input signals are mapped as $W$8, readout is formed from a subset of nodes, and the final output is $W$9, with $f$0 trained by ridge regression using Tikhonov parameter $f$1. The key empirical result is that intermediate network density gives the richest and most diverse nonlinear readouts, whereas dense networks distribute voltage so uniformly that memristive evolution is suppressed. On Lorenz63 autonomous prediction, a 500-node sparse network with 2119 edges achieves about 4.9 Lyapunov times, an intermediate-density 9905-edge network achieves about 6.3 Lyapunov times, and a very dense 123,671-edge network collapses to only 0.7 Lyapunov times (Xu et al., 22 May 2025). The same paper reports active-edge fractions of about $f$2 in the sparsest case, roughly $f$3 in the medium-density case, and only about $f$4 in the densest case, linking prediction quality directly to internal physical activity.

Memristor-based reservoirs provide a different physical DyRC mechanism: controlled heterogeneity of device dynamics. Ion-channel-based memristors are tuned by alamethicin concentration and DC offset so that distinct devices respond differently to the same encoded waveform, allowing single-encoding reservoir computing without reliance on masking or stochastic device-to-device variability. In a second-order nonlinear dynamical system prediction task, a varied memristor reservoir experimentally achieves a testing NMSE of $f$5 using only five distinct memristors, and in a neural activity classification task a reservoir of three distinct memristors experimentally attains $f$6 accuracy with a CNN readout (Armendarez et al., 2023). Here the DyRC content lies in the deliberate engineering of dynamic diversity rather than in network connectivity.

Other substrates show that DyRC is not confined to nano-electronics. “Road traffic reservoir computing” uses traffic flow itself as the reservoir, with signal phases, vehicle interactions, and road-network dynamics supplying memory and nonlinear coupling. On the Tsukuba temperature task, the paper reports that a 150-minute-ahead forecast is feasible with moderate accuracy, while internal traffic-density prediction remains possible even when only part of the network is used (Ando et al., 2019). Active colloidal oscillators realize a fully parallel physical reservoir of 400 hydrodynamically coupled oscillators with tunable coupling strength and fading-memory time. For Mackey–Glass forecasting, the reported example reaches NRMSE $f$7, and for hidden anomalies that preserve instantaneous statistical properties but disrupt temporal correlations the experimental reservoir achieves $f$8-score $f$9 (Heuthe et al., 9 Jan 2026). Coupled MEMS drum resonators implement yet another DyRC-like platform, where sideband-pumped phonon-cavity electromechanics, pump amplitude modulation, and time-delay feedback generate the reservoir states; the paper reports near-perfect parity prediction for low orders in optimal modulation windows and an electrical energy per input of about $$\mathbf{x}(t+\Delta t)=M(t)\,W_{\rm in}\mathbf{u}(t)$$0 in the SiN-probing case (Farah et al., 6 Jan 2026).

5. Attractor reconstruction, bifurcation learning, crises, and unseen state space

A distinguishing ambition of DyRC is to learn dynamical organization beyond local trajectory fitting. In a minimal-data digital-twin study of the externally forced Duffing oscillator, an auto-regressive echo-state network is trained on only two trajectories of length $$\mathbf{x}(t+\Delta t)=M(t)\,W_{\rm in}\mathbf{u}(t)$$1 at forcing amplitudes $$\mathbf{x}(t+\Delta t)=M(t)\,W_{\rm in}\mathbf{u}(t)$$2 and $$\mathbf{x}(t+\Delta t)=M(t)\,W_{\rm in}\mathbf{u}(t)$$3, corresponding to period-2 and period-4 dynamics. The trained model then predicts period-1 dynamics, period-3 dynamics, chaotic trajectories, and approximate bifurcation points across the full validation range $$\mathbf{x}(t+\Delta t)=M(t)\,W_{\rm in}\mathbf{u}(t)$$4. The same framework extends to a two-parameter $$\mathbf{x}(t+\Delta t)=M(t)\,W_{\rm in}\mathbf{u}(t)$$5 setting using only four training trajectories at $$\mathbf{x}(t+\Delta t)=M(t)\,W_{\rm in}\mathbf{u}(t)$$6 and $$\mathbf{x}(t+\Delta t)=M(t)\,W_{\rm in}\mathbf{u}(t)$$7, reproducing the shift of bifurcation structure with frequency (Yadav et al., 2024). The explicit inclusion of the forcing signal in the input makes the RC parameter-aware in the sense that parameter changes are part of the dynamical model rather than external metadata.

Crisis prediction makes the same point more sharply. In a parameter-aware reservoir with the bifurcation parameter injected as an input channel, the trained autonomous network is analyzed dynamically after training on data from parameter values before the crisis. For the logistic map, the trained reservoir predicts the boundary crisis at approximately $$\mathbf{x}(t+\Delta t)=M(t)\,W_{\rm in}\mathbf{u}(t)$$8, very close to the true $$\mathbf{x}(t+\Delta t)=M(t)\,W_{\rm in}\mathbf{u}(t)$$9, and fixed-point analysis identifies seven fixed points, of which two are stable. More importantly, the reservoir reproduces the same crisis skeleton as the original nonlinear system: unstable fixed points align with crisis-organizing structures, the bifurcation diagram shows attractor collision and disappearance, and the scaling law $M(t)=M\!\big(t;\mathbf{x}(t);\mathbf{u}(t)\big)$0 with $M(t)=M\!\big(t;\mathbf{x}(t);\mathbf{u}(t)\big)$1 is recovered near the boundary crisis (Sisodia et al., 15 Oct 2025). The paper also reports analogous findings for the Gauss map and notes that higher-dimensional examples such as the Hénon map are harder to visualize mechanistically.

DyRC-style generalization can also target disconnected regions of state space. A multiple-trajectory training scheme for reservoir computing stacks disjoint short trajectories into a single ridge-regression problem, allowing training on collections of transient-rich signals rather than one long orbit. Applied to the Duffing system and a magnetic pendulum, the study shows that RCs trained on trajectories from a single basin of attraction can generalize to unobserved basins. In the magnetic-pendulum setting, one reported configuration achieves $M(t)=M\!\big(t;\mathbf{x}(t);\mathbf{u}(t)\big)$2, compared with a baseline near $M(t)=M\!\big(t;\mathbf{x}(t);\mathbf{u}(t)\big)$3, while spurious attractors are rare at $M(t)=M\!\big(t;\mathbf{x}(t);\mathbf{u}(t)\big)$4 (Norton et al., 5 Jun 2025). The paper explicitly argues that this occurs without explicit structural priors, suggesting that the reservoir projection and ridge-regression readout can act as an implicit inductive bias for dynamical extrapolation.

6. Misconceptions, limitations, and open problems

A common misconception is that success in reservoir forecasting is adequately characterized by short-horizon error. Several DyRC papers argue otherwise. The Hénon-attractor study explicitly shows that good trajectory prediction does not automatically imply correct geometric reconstruction of the attractor: as spectral radius increases, the reservoir attractor becomes “fatter,” the box-counting dimension rises above the true value, and the unrestricted reservoir Lyapunov spectrum becomes harder to interpret directly even though the inertial manifold may still carry the correct dynamics (Kobayashi et al., 16 Sep 2025). Similarly, crisis-prediction work emphasizes that the point is not merely that the reservoir forecasts a crisis, but that its internal autonomous dynamics reconstruct the same fixed points, unstable branches, and return-map geometry as the source system (Sisodia et al., 15 Oct 2025).

A second misconception is that stronger recurrence or denser connectivity is automatically beneficial. The nanowire-network study directly contradicts this: too little connectivity limits signal propagation and mixing, while too much connectivity suppresses the voltage heterogeneity needed to activate memristive dynamics. Performance therefore peaks at intermediate density rather than maximal density (Xu et al., 22 May 2025). DyRC-VG reaches a related conclusion from a different angle: network metrics such as clustering coefficient and density appear relevant, but the paper does not claim a definitive causal conclusion yet, indicating that structure-function relations in reservoirs remain only partially understood (Geier et al., 25 Jul 2025).

Physical DyRC platforms also retain practical constraints. Memristor reservoirs with deliberately diverse dynamics still rely on careful tuning of concentration and bias, use no inter-node coupling in the reported parallel configuration, and face scaling challenges associated with fabricated droplet interface bilayers (Armendarez et al., 2023). Active colloidal reservoirs require optical steering, microscopy feedback, and Gaussian-kernel feature extraction, while coupled MEMS drum reservoirs show that multimode dynamics can be exploited but are not yet superior to the best single-resonator kHz MEMS reservoirs on NARMA benchmarks (Heuthe et al., 9 Jan 2026, Farah et al., 6 Jan 2026). In addition, the interpretability that DyRC seeks is strongest in low-dimensional systems; the crisis-prediction study notes that in a 400-dimensional reservoir trained on the Hénon map, stable and unstable manifolds are hard to reconstruct and the collision mechanism is much harder to visualize (Sisodia et al., 15 Oct 2025).

At the same time, the literature does not support an opposite simplification that explicit physics priors are always necessary. The unseen-state-space study shows that reservoir computers can generalize to entirely unobserved basins without explicit structural priors when training data sufficiently sample informative transients (Norton et al., 5 Jun 2025). This suggests that DyRC spans a continuum from explicit structure injection to implicit dynamical regularization. The unresolved question is not whether dynamics matter—they do across all of these formulations—but which aspects of dynamics should be encoded in the reservoir, the readout, or the training protocol for a given class of systems.

DyRC therefore marks a shift in the interpretation of reservoir computing. Rather than viewing the reservoir as an arbitrary high-dimensional helper for linear regression, DyRC treats it as a model of dynamical mediation: a synchronization manifold, a multiscale memory kernel, a data-derived graph, or a physical medium whose internal laws shape representation and prediction. The field’s most developed results indicate that this shift is useful precisely when one seeks not only forecasts, but attractors, Lyapunov structure, bifurcations, crises, unseen basins, and physically meaningful computation (Xu et al., 22 May 2025, Kobayashi et al., 16 Sep 2025, Yadav et al., 2024).