Recurrent neural chemical reaction networks trained to switch dynamical behaviours through learned bifurcations

Both natural and synthetic chemical systems not only exhibit a range of non-trivial dynamics, but also transition between qualitatively different dynamical behaviours as environmental parameters change. Such transitions are called bifurcations. Here, we show that recurrent neural chemical reaction networks (RNCRNs), a class of chemical reaction networks based on recurrent artificial neural networks that can be trained to reproduce a given dynamical behaviour, can also be trained to exhibit bifurcations. First, we show that RNCRNs can inherit some bifurcations defined by smooth ordinary differential equations (ODEs). Second, we demonstrate that the RNCRN can be trained to infer bifurcations that allow it to approximate different target behaviours within different regions of parameter space, without explicitly providing the bifurcation itself in the training. These behaviours can be specified using target ODEs that are discontinuous with respect to the parameters, or even simply by specifying certain desired dynamical features in certain regions of the parameter space. To achieve the latter, we introduce an ODE-free algorithm for training the RNCRN to display designer oscillations, such as a heart-shaped limit cycle or two coexisting limit cycles.

💡 Research Summary

The paper investigates how recurrent neural chemical reaction networks (RNCRNs), a class of chemical reaction networks (CRNs) that emulate recurrent artificial neural networks, can be trained not only to reproduce a prescribed dynamical behavior but also to exhibit bifurcations—qualitative changes in dynamics triggered by smooth variations of environmental parameters. The authors first extend the original RNCRN framework by introducing “parameter species” (Λₗ) whose concentrations encode the values of external parameters. These species are static (dλₗ/dt = 0) but interact catalytically with the fast chemical perceptrons (Yⱼ), thereby allowing the perceptrons’ activation functions to depend on the parameters in a smooth way. This construction makes it possible to train a single RNCRN to approximate a family of parameter‑dependent ODEs f(x, λ) over a range of λ values, using essentially the same learning algorithm that was previously applied to parameter‑free systems.

To demonstrate the capability of the parametrized RNCRN, the authors train it on two classic bifurcations. In the Hopf example, they use a shifted normal‑form system with an equilibrium at (5, 5) and a critical parameter λ* = 2. Training on data taken only at λ ≈ 2, the RNCRN automatically reproduces the transition from a stable fixed point (λ < 2) to a stable limit cycle (λ > 2). In the homoclinic case, a similar procedure yields a network that switches from a fixed point to a homoclinic orbit as the parameter crosses its critical value. These results show that the RNCRN can inherit smooth bifurcation structures without being explicitly supplied with bifurcation information.

The paper then tackles a more challenging scenario: target dynamics that are discontinuous with respect to the parameters, or even defined only by sparse samples of a desired limit‑cycle shape. For this, the authors devise an ODE‑free training algorithm. Instead of providing an explicit vector field, they supply a set of points sampled from the desired limit cycle (e.g., a heart‑shaped orbit or two coexisting cycles). The loss combines a Chamfer‑type distance between the network‑generated trajectory and the sampled points with penalties for mismatched period and phase. Remarkably, with only a few hundred sampled points, the RNCRN learns a full vector field that generates the prescribed oscillation, correctly reproducing its geometry, period, and phase. This demonstrates that RNCRNs can extrapolate from sparse data to construct coherent dynamical systems.

Further, the authors show that when the target system is piecewise‑defined—different ODEs apply in different regions of parameter space—the RNCRN can learn to “detect” the region via the parameter species and automatically switch to the appropriate dynamics, despite never being shown an explicit bifurcation diagram. This conditional behavior emerges from the nonlinear catalytic influence of the Λₗ species on the perceptrons, effectively implementing a learned decision boundary inside the chemical network.

Finally, the paper integrates a simple chemical classifier with the RNCRN, enabling the network to sense an external signal, perform a nonlinear classification, and then toggle between distinct dynamical regimes (e.g., from a quiescent state to an oscillatory one). This illustrates a potential route toward synthetic biological circuits that combine sensing, decision‑making, and dynamic control using purely chemical reactions.

In summary, the work makes three major contributions: (1) a generalized RNCRN architecture that incorporates static parameter species, allowing the network to learn smooth, parameter‑dependent vector fields and associated bifurcations; (2) empirical validation that the network can reproduce classical Hopf and homoclinic bifurcations and can learn discontinuous, piecewise dynamics without explicit bifurcation data; and (3) an ODE‑free training methodology that enables the design of complex limit‑cycle shapes from sparse trajectory samples. These advances broaden the applicability of chemical‑reaction‑based neural computation to synthetic biology, DNA nanotechnology, and chemical computing, especially in contexts where only limited or discontinuous dynamical specifications are available.