betaFeedback

WorksPhysics and Astronomy

Delay-induced current reversal in traveling activity waves

Abstract

Self-propelled particles that regulate their speed based on a response to a local activity source can move with or against a traveling wave of that source depending on their response time. Silica-titania Janus microswimmers, whose UV-controlled speed relaxes with a delay intrinsic to the titania photoconductivity adaptation, carry two independent memory channels: the familiar directional memory dictated by persistent active Brownian motion, and a distinct memory from the speed adaptation. Combining experiments, Langevin simulations, and a moment expansion of the corresponding Fokker-Planck equation, we show that this delayed response to speed adaptation reshapes transport under traveling UV-illumination waves and can even reverse it relative to the direction set by orientational memory alone. In particular, we observe that delay promotes transport in the direction opposite to the traveling activity waves, where, beyond a critical value, the particle current reverses with respect to the wave propagation direction. Crucially, this behavior requires an asymmetric response to the speed adaptation, with an instantaneous rise but a delayed fall upon exposure to time-varying UV light. Theory further shows that this phenomenology is specific to the response asymmetry rather than to the mere presence of a delay, i.e. reversing the asymmetry reverses the effect of the delay on the particle current relative to the wave. We thus establish propulsion-delay asymmetry as a control parameter for directed transport of responsive active particles in traveling activity landscapes.

A finite response time is central to how bacteria sense and navigate chemical gradients in their environment. While the concentration difference across a single cell’s own length is small, a cell can probe a far larger stretch of the gradient than its body alone would allow by comparing concentrations sensed at different times as it swims [1, 2].

More generally, when a microswimmer, biological or synthetic, adjusts its motion in response to a time-varying activity cue, such as a traveling chemical wave, a finite response time can give rise to net transport up or downstream the traveling pattern. The direction and magnitude of this transport depend on a competition between timescales: on one side, the wave period τℓ=ℓ/c\tau_{\ell} = \ell/c, set by the wavelength ℓ\ell and the wave speed cc, and on the other side, the timescales controlling the microswimmer motion, namely the swim time τu=ℓ/umax⁡\tau_{u} = \ell/u_{\max} needed to cross one wavelength at its maximum speed umax⁡u_{\max}, and the internal memory times over which the microswimmer retains information about its past, here its orientational persistence and the delay in adapting its propulsion.

For instance, in the chemotactic amoeba Dictyostelium discoideum, the relevant memory is the decay time of the chemotactic response. When it matches the wave period, cells are desensitized to the trailing edge of a self-generated, outward-traveling chemoattractant wave [3], which rectifies their motion toward the source and allows them to aggregate [5] [4].

For active Brownian particles whose speed adapts instantaneously, an orientational memory is anyway present and determined by rotational diffusion: the particles self-propel with persistent motion with a rotational decorrelation time τr≡1/Dr\tau_{r} \equiv1/D_{r}, where DrD_{r} is the rotational diffusion coefficient. With no deterministic coupling to any local gradient, orientational memory alone drives transport against slow waves and along fast ones, with τr\tau_{r} setting the wave speed at which the effect peaks [6, 7, 8, 9]. Alternatively, the presence of a deterministic aligning torque can lead to the same transport reversal by a different route: here, memory emerges in the form of alignment time, which competes with the time spent inside a moving activity pulse, as demonstrated for Janus colloids whose orientation aligns toward a light gradient [10].

In all these cases, the memory resides in the orientation or, for Dictyostelium, in a chemotactic response. A self-propelled particle can, however, carry a second, independent memory: a delay τ\tau in the adaptation of its own propulsion speed in response to a local variation of activity. In most synthetic swimmers with adaptive propulsion this delay is small and typically neglected [11, 12]. In some systems, however, it is large enough to matter, as for genetically engineered photokinetic bacteria [13, 14] and for photoconductive Janus particles, where speed adaptation to light modulation is an intrinsic material response [15]. The two memories differ in kind. Orientational persistence locks in the particle’s own past direction up to τr\tau_r and is blind to an activity wave, whereas a propulsion delay records the wave’s recent state, so that propulsion reflects where the wave was rather than where the particle is [16]. Imposed delays are known to induce and reverse net currents in passive Brownian ratchets [17, 18] and have been engineered into active agents through feedback loops [19, 20, 21, 22, 23, 24]. Delayed speed adaptation, including exponential relaxation, was further predicted to rectify transport in spatially asymmetric, periodic activity landscapes and to reverse it upon tuning the delay [25]. All these studies, however, consider static or feedback-generated activity landscapes [22, 23, 25], not an externally imposed traveling wave, which the particle intrinsically responds to.

In this Letter, combining experiments, Langevin simulations, and a moment expansion of the Fokker-Planck equation, we show that delay biases transport against the wave and that, beyond a critical value, reverses the particle current. We also show that current reversal requires an asymmetric time response.

Experimental system and simulations

We use silica-titania Janus microswimmers, 2 μ\mum in diameter, suspended in water with Pluronic F-127 and sedimented onto the lower of two ITO electrodes confining a thin liquid cell. A uniform AC electric field propels them by induced-charge electrophoresis [26, 27], while the photoconductive titania cap lets UV illumination modulate the propulsion speed orthogonally to the driving field [15]. Projecting UV patterns into the cell with a digital micromirror device (DMD) thus lets us impose spatiotemporal activity landscapes on the swimmers.

In these systems the propulsion speed vv grows with light intensity up to a saturation level, Figure 1(a), as previously reported for photocatalytic microswimmers [28, 29, 30]. Propulsion does not, however, respond instantaneously to the removal of illumination. It relaxes toward its target value with an intrinsic delay set by the photoconductivity of the titania and the lifetime of its charge carriers [15, 31, 32]. Switching illumination on and off directly reveals this memory in speed: switching the UV on, the active speed rises essentially instantaneously (see SM [33], Figure S1); switching it off, it relaxes smoothly, following a single exponential decay with a time constant τ≈2\tau\approx2 s, Figure 1(b). This asymmetric response time adds on to a timescale characteristic of the decorrelation of the particle orientation set by Brownian rotational diffusion, τr≈6.3\tau_r \approx6.3 s. The DMD moves the illumination pattern as a traveling wave, Figure 1(c,Figure 1d). As the particles are illuminated, they adjust their propulsion speed to a local “target speed” set by the local light intensity and retain it over a time ∼τ\sim\tau once the intensity decreases.

Figure 1. Propulsion response of the Janus microswimmers to UV illumination. (a) Active speed vs. light intensity, normalized by the maximum speed, from MSD fits across a range of illumination intensities; error bars are standard deviations. Dashed line: fit to a saturating hyperbolic response. Data redrawn from Ref. [15]. (b) Active speed following a spatially uniform switch-off of UV illumination, normalized and averaged over three illumination cycles across four experimental cells; shaded regions are 95% confidence intervals. Dashed line: fitted exponential decay, giving a delay of 2 s. (c) Micrograph of microswimmers under the traveling-wave pattern at steady state (umax⁡=8 μm s−1u_{\max}=8\,\mu\mathrm{m}\,\mathrm{s}^{-1}, c=0.6 μm s−1c=0.6\,\mu\mathrm{m}\,\mathrm{s}^{-1}), showing particle trajectory tails over the preceding 5 s, color-coded by instantaneous speed, with a sketch below showing the wave intensity profile. (d) Schematic of the setup: the particle suspension is confined between two ITO-coated glass slides, contacted with copper tape to apply the AC field, and a DMD projects the light pattern through the objective into the sample plane.

Figure S1. Switch-on response of the active speed. Active speed following a spatially uniform switch-on of UV illumination, normalized and averaged over three illumination cycles across four experimental cells; shaded regions are 95% confidence intervals. The rise reaches steady state within 0.4 s, close to the tracking’s temporal resolution.

Correspondingly, we model the system as overdamped active Brownian particles with the same asymmetric, first-order propulsion lag, driven by a sinusoidal traveling wave (simulation details in the Supplemental Material [33]).

Model. In the frame comoving with the wave, z=x−ctz=x-ct, with xx the lab-frame position, the particle’s Langevin dynamics for orientation θ\theta and active speed vv reads

z˙=vcos⁡θ−c+2D η,\dot{z}=v\cos\theta-c+\sqrt{2D}\,\eta,
θ˙=2Dr χ,\dot{\theta}=\sqrt{2D_r}\,\chi,
v˙=−τ(z,v)−1[v−u(z)],(1)\dot{v}=-\tau(z,v)^{-1}[v-u(z)], \tag*{(1)}

with η,χ\eta,\chi independent Gaussian white noises, D,DrD,D_r the translational and rotational diffusion coefficients, u(z)u(z) the traveling pattern’s target speed, and τ(z,v)\tau(z,v) the response time measured above: an instantaneous rise (τon≈0\tau_{\mathrm{on}}\approx0) and a delayed fall (τoff=τ\tau_{\mathrm{off}}=\tau). When τ=0\tau=0, v=u(z)v=u(z) exactly and the standard active Brownian particle model is recovered. A finite τ\tau is what lets vv carry memory of the particle’s own history. The Fokker-Planck equation for the joint probability density ψ(z,θ,v,t)≡⟨δ(z−zt)δ(θ−θt)δ(v−vt)⟩\psi(z,\theta,v,t)\equiv\langle\delta(z-z_t)\delta(\theta-\theta_t)\delta(v-v_t)\rangle follows directly from these Langevin equations:

∂tψ=−∂z[(vcos⁡θ−c)ψ]+D∂z2ψ+Dr∂θ2ψ+∂v[(v−u(z)τ(z,v))ψ].(2)\partial_t\psi=-\partial_z\left[(v\cos\theta-c)\psi\right]+D\partial_z^2\psi+D_r\partial_\theta^2\psi+\partial_v\left[\left(\frac{v-u(z)}{\tau(z,v)}\right)\psi\right]. \tag*{(2)}

The first three terms in Eq. (2) are advection along zz at the particle’s own speed relative to the wave, translational diffusion, and rotational diffusion of the orientation; the fourth is the relaxation of the active speed toward the local wave target just introduced above. We seek steady-state solutions of the equation that preserve the symmetry θ→−θ\theta\to-\theta, which allows us to expand ψ\psi in cosine harmonics, Ψk≡∫cos⁡(kθ) ψ dθ\Psi_k\equiv\int\cos(k\theta)\,\psi\,d\theta. This reduces the problem to a hierarchy of moment equations, which we close at second order (Ψ3≡0\Psi_3\equiv0), retaining the particle number density (Ψ0\Psi_0), the polar order (Ψ1\Psi_1), and the nematic order (Ψ2\Psi_2). Its root mean square amplitude grows with increasing persistence length in the simulations (Figure S3), and the second-order closure reproduces this trend with good agreement, consistent with general studies of active Brownian particle moment closures [34] (the full scaling argument is given in the Supplemental Material [33]). We solve the resulting closed linear system numerically as a sparse linear system without time-stepping (full scheme in the Supplemental Material [33]), and the solution is then nondimensionalized in space for comparison with experiments and simulations, with ζ≡z/ℓ\zeta\equiv z/\ell the resulting dimensionless, periodic phase. This gives us direct access to the phase-resolved density ρ(ζ)≡ℓ∫Ψ0 dv\rho(\zeta)\equiv\ell\int\Psi_0\,dv, and, from the polar-order flux q(ζ)≡ℓ∫vΨ1 dvq(\zeta)\equiv\ell\int v\Psi_1\,dv, the local mean velocity vζ=q(ζ)/ρ(ζ)v_\zeta=q(\zeta)/\rho(\zeta) along the wave at phase ζ\zeta, without free parameters beyond those fixed by the experiments or simulations.

Figure S3. Nematic order requires closing the hierarchy at Ψ3≡0\Psi_3 \equiv0. (a) Root-mean-square (RMS) nematic order ⟨cos⁡2ϑ⟩(ζ)\langle\cos2\vartheta\rangle(\zeta) measured directly in simulated trajectories (horizontal axis) compared against the Ψ3≡0\Psi_3 \equiv0 model’s own prediction (vertical axis; the Ψ2≡0\Psi_2 \equiv0 closure predicts this identically zero and so has no prediction to show here), for every tested combination of umax⁡u_{\max}, τ∈{0,2,4,8} s\tau\in\{0, 2, 4, 8\}\,\mathrm{s}, and c∈{2,4,6,8,10} μm s−1c \in\{2, 4, 6, 8, 10\}\,\mathrm{\mu m\,s^{-1}}, colored by the persistence-length ratio ℓp/ℓ≡umax⁡/(Drℓ)\ell_p/\ell\equiv u_{\max}/(D_r\ell). Dashed line: y=xy=x. Agreement confirms the model correctly predicts the amplitude of the real nematic signal. (b) Root-mean-square nematic order predicted by the Ψ3≡0\Psi_3 \equiv0 model (color), swept over persistence length ℓp≡umax⁡/Dr\ell_p \equiv u_{\max}/D_r and wavelength ℓ\ell independently, at fixed particle radius R=1.03 μmR=1.03\,\mathrm{\mu m} (fixing DrD_r) and wave speed c=10 μm s−1c=10\,\mathrm{\mu m\,s^{-1}}, in the zero-delay limit.

Results

Figure 2 shows the phase-resolved density and local mean velocity from experiments (panel a, markers), simulations (panels b, c, markers), and theory (solid lines). For the experimental panels, the theory uses the illumination profile u(ζ)u(\zeta) measured directly from the data, rather than the idealized sinusoid prescribed by the DMD. This measured profile automatically captures the particles’ actual speed response, including the nonlinear relation between light intensity and active speed; its derivation, together with how the instantaneous velocity profiles are debiased for the influence of diffusion using a Rice distribution, is given in the Supplemental Material [33]. The theory reproduces both quantities closely across the full range of tested wave speeds and delays. Both experiments and simulations fall also close to the predicted peak and trough values in parity plots given in the Supplemental Material (Figure S4) [33].

Figure 2. Density and local mean velocity: experiment, simulation, and theory. Top row, ρ(ζ)\rho(\zeta); bottom row, vζv_\zeta. Markers: experimental (a) or simulated (b, c) data; solid lines: moment expansion. (a) Experiments and theory for different wave speed cc (color bar), for umax⁡u_{\max} between 6 μm s−1\mu\mathrm{m}\,\mathrm{s}^{-1} to 8 μm s−1\mu\mathrm{m}\,\mathrm{s}^{-1}. (b) Simulation at fixed umax⁡=5 μm s−1u_{\max}=5\,\mu\mathrm{m}\,\mathrm{s}^{-1}, τ=2 s\tau=2\,\mathrm{s}, for different cc (color bar). (c) Simulation at fixed umax⁡=5 μm s−1u_{\max}=5\,\mu\mathrm{m}\,\mathrm{s}^{-1}, c=5 μm s−1c=5\,\mu\mathrm{m}\,\mathrm{s}^{-1}, for different delay τ\tau (color bar).

Figure S4. Theory–experiment and theory–simulation parity for the local mean velocity. Predicted (theory) vs. measured or simulated peak and trough of vζ(ζ)v_{\zeta}(\zeta), one peak and one trough per run; both are extracted by fitting a periodic smoothing spline to the phase-binned curve and reading off its extrema, applied identically to the theory’s noise-free solution for consistency. Circles: peak; squares: trough; color: wave speed cc, 1 µm s−1^{-1} to 10 µm s−1^{-1}. Dashed line: y=xy=x. (a) Experiment, all runs, using each run’s own measured wave speed and particle speed, fixed DD and DrD_{\mathrm{r}}, fixed delay τ=2\tau=2 s, and the empirically measured driving profile (Driving profile measurement, above). (b) Simulation, a representative subset spanning umax⁡∈{1,3,5,8}u_{\max} \in\{1,3,5,8\} µm s−1^{-1}, τ∈{0,2,4,8}\tau\in\{0,2,4,8\} s, and c∈{1,3,5,8,10}c \in\{1,3,5,8,10\} µm s−1^{-1}.

At low wave speed, density peaks just before ζ=1\zeta=1, where the driving profile itself vanishes (u(ζ)→0u(\zeta)\to0), as particles slow toward this point and linger. This is the classic result that active particles in a spatially-varying speed field accumulate in slow regions, with density ρ(r)∝1/v(r)\rho(r)\propto1/v(r) [35, 36], exploited directly for density shaping in photokinetic bacteria [13] and in our own silicatinia microswimmers [15]. This result is also recovered analytically as the low-wave-speed limit of the same traveling-activity-wave problem [6]. As wave speed grows and approaches the particles’ own speed, this quasi-static picture breaks down and the peak weakens, lowering and shifting away from ζ=1\zeta=1 while the trough shallows and shifts further behind. This is the same crossover toward a flattened, near-uniform density predicted by the same theory at high wave speed [6]. The density peak has little sensitivity to delay at fixed wave and particle speed, since its position is set almost entirely by how much the particles’ own speed exceeds the wave speed. This follows directly from the asymmetric relaxation itself. The peak forms as particles enter the profile under the fixed, near-instantaneous τon\tau_{\mathrm{on}}, so it changes little as τoff=τ\tau_{\mathrm{off}} = \tau is swept, while only the trailing edge, governed by τoff\tau_{\mathrm{off}}, broadens as the delay grows.

The local mean velocity vζv_{\zeta} shows the same two-sided structure in both experiment and simulation, with a positive peak just ahead of the wave crest, driven forward by the wave, and a negative trough behind it, where particles fall back. As wave speed increases, the peak first rises, then falls again while shifting toward the crest, and the trough shallows while shifting further behind, mirroring the density profile’s own evolution. Delay leaves the peak essentially unchanged, consistent with the same τon\tau_{\mathrm{on}}-governed insensitivity seen in the density, but deepens the trough at first, saturating before it broadens further with increasing τ\tau, on the τoff\tau_{\mathrm{off}}-governed side of the profile.

The quantity of central interest is the net drift velocity, i.e., the net displacement per unit time that the traveling wave imparts on the particles, obtained as the density-weighted average of vζv_{\zeta} over one period,

vtransport=∫01ρ(ζ) vζ dζ.(3)v_{\mathrm{transport}} = \int_{0}^{1} \rho(\zeta)\,v_{\zeta}\,\mathrm{d}\zeta. \tag*{(3)}

We take vtransport>0v_{\mathrm{transport}} > 0 for a net particle current along the direction of wave propagation and vtransport<0v_{\mathrm{transport}} < 0 for net transport against it. Equivalently, vtransport=ℓJv_{\mathrm{transport}} = \ell J, with JJ the stationary probability current [6], so that a sign change of vtransportv_{\mathrm{transport}} constitutes a current reversal.

Figure 3(a) shows vtransportv_{\mathrm{transport}} vs. wave speed. Experimental data, overlaid with the theory at fixed τ=2 s\tau= 2\,\mathrm{s}, agree qualitatively, with the direction of net transport reversing with wave speed, running against a slow wave and along with a fast wave, respectively. This is the same crossover already established for orientational memory alone, without any delay in speed [6, 7, 8, 9]. The role of the delay becomes clear once τ\tau is varied at fixed particle speed in the simulations, Figure 3(b). The curve’s overall shape, and the wave speeds at which it reaches its maximum and minimum, change little with τ\tau. Instead, increasing τ\tau shifts the whole curve downward, more strongly at its maximum than at its minimum. The shift continues until, for sufficiently large τ\tau, transport with the wave is eliminated entirely, leading to a net current against the wave across the full range tested. This behavior follows directly from the phase-resolved behavior established above: delay leaves the leading part of the ρ\rho and vζv_{\zeta} profiles largely unaffected while deepening their trailing part, shifting the density-weighted net transport against the wave as τ\tau grows.

Figure 3. Net drift velocity vtransportv_{\mathrm{transport}} vs. wave speed cc. Markers: experimental (a) or simulated (b) data; solid line: moment expansion. (a) Experimental vtransportv_{\mathrm{transport}} vs. cc, colored by particle speed umax⁡u_{\max}, at fixed τ=2 s\tau= 2\,\mathrm{s}. (b) Simulated vtransportv_{\mathrm{transport}} vs. cc, colored by delay τ\tau, at fixed umax⁡=8 μm s−1u_{\max} = 8\,\mu\mathrm{m}\,\mathrm{s}^{-1}.

Besides the timescales introduced above, there is also the translational-diffusion time across one wavelength, τD≡ℓ2/D\tau_D \equiv\ell^2/D. For a given profile shape u(ζ)/umax⁡u(\zeta)/u_{\max}, the transport is thus fully set by umax⁡/c=τℓ/τuu_{\max}/c = \tau_{\ell}/\tau_u and the ratios τ/τℓ\tau/\tau_{\ell}, τD/τℓ\tau_D/\tau_{\ell}, τD/τr\tau_D/\tau_r. The latter ratio depends only on particle size and wavelength, which are fixed in our simulations and experiments. Figure 4 therefore maps the net transport across the (umax⁡/c)−(τ/τℓ)(u_{\max}/c)-(\tau/\tau_{\ell}) plane, beyond the wave speeds and delays tested experimentally, isolating the role of the delay’s asymmetry. Here τ/τℓ\tau/\tau_{\ell} measures how large a fraction of one wave-period the delay spans. Meanwhile, τD/τℓ>103\tau_D/\tau_{\ell} > 10^3 across the entire range, so translational diffusion never competes with the wave-crossing timescale. Our experimental setup itself restricts τ/τℓ\tau/\tau_{\ell} to below ≈0.1\approx0.1. The used wavelength ℓ=200 μm\ell= 200\,\mu\mathrm{m} balances the DMD dithering resolution against fitting several wavelengths within the field of view, and cc is limited by the DMD’s ≈50 ms\approx50\,\mathrm{ms} update time needed to move the activity wave smoothly. umax⁡/cu_{\max}/c, by contrast, reaches up to ≈50\approx50 in the experiments.

Figure 4. Net drift velocity for asymmetric, symmetric, and reversed-asymmetric delay. Theoretical net drift velocity vtransportv_{\mathrm{transport}} (color, μm s−1\mu\mathrm{m}\ \mathrm{s}^{-1}, color map centered at zero), solved directly from the moment expansion of Eq. (2) at fixed particle radius R=1.03 μmR = 1.03\ \mu\mathrm{m} (giving τD/τr≈5.66×104\tau_D/\tau_r \approx5.66 \times10^4) and wavelength ℓ=200 μm\ell= 200\ \mu\mathrm{m}. (a) Asymmetric delay (τon≈0\tau_{\mathrm{on}} \approx0, τoff=τ\tau_{\mathrm{off}} = \tau), as used throughout this work; reproduces the same regime as Figure 3. (b) Symmetric delay (τon=τoff=τ\tau_{\mathrm{on}} = \tau_{\mathrm{off}} = \tau). (c) Reversed asymmetric delay (τon=τ\tau_{\mathrm{on}} = \tau, τoff≈0\tau_{\mathrm{off}} \approx0).

With the asymmetric delay used throughout this work, Figure 4(a), transport with the wave survives only in a narrow region around umax⁡/c≈1u_{\max}/c \approx1 and weakens as umax⁡/cu_{\max}/c moves away from this value in either direction, or as τ/τℓ\tau/\tau_{\ell} grows. Across most of the plane, particles are transported against the wave, a regime reached by increasing either umax⁡/cu_{\max}/c or τ/τℓ\tau/\tau_{\ell}, so delay alone can reverse the particle flow with respect to the wave propagation direction. Making the delay symmetric instead, Figure 4(b), removes this ability entirely: increasing τ/τℓ\tau/\tau_{\ell} only ever weakens transport in whichever direction is already set by umax⁡/cu_{\max}/c, so delay becomes uniformly detrimental to transport magnitude, and reversal is possible only by varying umax⁡/cu_{\max}/c, never the delay itself. Conversely, reversing the asymmetry, Figure 4(c), flips the picture entirely: the system is now dominated by transport with the wave, which is enhanced by delay rather than suppressed, with transport against the wave confined to a small region at umax⁡/c≳3u_{\max}/c \gtrsim3 and very small delays. Together, the three panels show that delay does not affect transport indiscriminately. It becomes a tuning parameter whose effect depends on the details of its functional form, including its sign and symmetry.

Conclusion. We have shown that silica-titania Janus microswimmers carry an additional distinct memory channel in speed adaptation beyond persistent active Brownian motion. This delay strongly reshapes the transport set by orientational memory alone in a traveling activity wave and can even reverse the particle flow with respect to the wave propagation direction. Analysis of the form of speed relaxation shows that the asymmetry between velocity rise and fall, not delay alone, is the genuine ingredient rectifying transport, in analogy to ratchets driven by temporally asymmetric zero-mean noise [37]. Reversing this asymmetry reverses its effect on transport entirely. This establishes the asymmetry of propulsion delay as a control parameter for directed transport that acts independently of the wave itself, and suggests that any system with an intrinsically delayed response to a traveling stimulus, biological or synthetic, may exploit the same competition between wave speed, delay, and persistence time.

U.T. thanks K. Studer, N. Windbacher, and R. Style for their time and for many good discussions. D.L. acknowledges A. Katona, S. Lévy, and I. Zurigel for insightful discussions. L.I. and U.T. acknowledge funding from the European Research Council (ERC) under the European Union’s Horizon 2020 Research and Innovation Program grant agreement No. 101001514. The authors also acknowledge the use of cleanroom facilities at the Binnig and Rohrer Nanotechnology Center (BRNC) at IBM Research – Zurich and at the FIRST Center for Micro- and Nanoscience at ETH Zurich. A.G. acknowledges AGAUR and FSE+ and Generalitat de Catalunya for financial support under the call FI SDUR 2023 Ref. CCI 2021ES05FPR011. M.C. and A.D. acknowledge financial support through Grant No. ProyExcel_00505, funded by Junta de Andalucía, and Grants No. PID2021-126348NB-I00 and No. PID2024-156257NB-C21, funded by MI-CIU/AEI/10.13039/501100011033 and by “ERDF/EU”. D.L. acknowledges MCIU/AEI for financial support under Projects No. PID2022-140407NB-C22 and No. PID2025-170781NB-C22. I.P. acknowledges support from Ministerio de Ciencia, Innovación y Universidades MCIU/AEI/FEDER under grant agreement PID2024-156516NB-10, and Generalitat de Catalunya under Program Icrea Acadèmia and project 2021SGR-673. Development of numerical solvers and simulation code was supported by Anthropic Claude Opus 5.5.

Data availability. The data are available from the authors upon reasonable request.

Acknowledgments

U.T. thanks K. Studer, N. Windbacher, and R. Style for their time and for many good discussions. D.L. acknowledges A. Katona, S. Lévy, and I. Zurigel for insightful discussions. L.I. and U.T. acknowledge funding from the European Research Council (ERC) under the European Union’s Horizon 2020 Research and Innovation Program grant agreement No. 101001514. The authors also acknowledge the use of cleanroom facilities at the Binnig and Rohrer Nanotechnology Center (BRNC) at IBM Research – Zurich and at the FIRST Center for Micro- and Nanoscience at ETH Zurich. A.G. acknowledges AGAUR and FSE+ and Generalitat de Catalunya for financial support under the call FI SDUR 2023 Ref. CCI 2021ES05FPR011. M.C. and A.D. acknowledge financial support through Grant No. ProyExcel_00505, funded by Junta de Andalucía, and Grants No. PID2021-126348NB-I00 and No. PID2024-156257NB-C21, funded by MI-CIU/AEI/10.13039/501100011033 and by “ERDF/EU”. D.L. acknowledges MCIU/AEI for financial support under Projects No. PID2022-140407NB-C22 and No. PID2025-170781NB-C22. I.P. acknowledges support from Ministerio de Ciencia, Innovación y Universidades MCIU/AEI/FEDER under grant agreement PID2024-156516NB-10, and Generalitat de Catalunya under Program Icrea Acadèmia and project 2021SGR-673. Development of numerical solvers and simulation code was supported by Anthropic Claude Opus 5.5.

Data availability. The data are available from the authors upon reasonable request.

References

Particle Fabrication

Silica–titania Janus particles were fabricated from monodisperse silica colloids (2.06 μ\mum diameter, 5 % w/v, microParticles GmbH, Germany), diluted to 0.5 % w/v in Milli-Q water. 100 μ\muL of the sonicated suspension were spread onto a glass slide cleaned by 1 min air-plasma treatment and allowed to dry into a colloidal monolayer. The monolayer was coated with 50 nm of Ti3_3O5_5 by electron-beam evaporation (Plassys MEB700SL) and subsequently annealed at 500 °C for 120 min under 1 bar O2_2 in a rapid thermal processor (AS-One 150, ANNEALSYS) to promote full oxidation and conversion to the TiO2_2 anatase phase. Particles were retrieved by sonication in Milli-Q water. Prior to each experiment, the dispersion was mixed in equal volumes with a 0.6 % w/v Pluronic F-127 solution, yielding a final surfactant concentration of 0.3 % w/v to reduce particle adhesion to the substrate.

Experimental cell

Observation cells were assembled from two indium–tin oxide coated glass electrodes (16 mm ×\times 16 mm, <7 Ω sq−1<7\,\Omega\,\mathrm{sq}^{-1}, präzisions glas & optik) sandwiched around a 120 μ\mum thick adhesive spacer with a 9 mm circular opening (Grace Bio-Laboratories SecureSeal, USA). The spacer was halved to provide an inlet and outlet, through which 9.5 μ\muL of particle suspension was introduced before sealing with vacuum grease (Krytox GPL205). Electrical contact was made with adhesive copper tape, and an AC electric field was applied via an arbitrary waveform generator (33522A, Agilent) at voltages between 1 V to 10 V and frequencies between 4 kHz to 10 kHz.

Microscopy and illumination

All experiments were performed on an inverted optical microscope (Eclipse Ti2-E, Nikon) using 20 ×\times (CFI S Plan Fluor ELWD 20XC) and 40 ×\times (CFI S Plan Fluor ELWD 40XC) objectives. UV illumination (λ=365\lambda= 365 nm, UHP-F-365, Prizmatix) was applied at intensities up to 6.3 W cm−2^{-2} and spatially patterned within the field of view using a DMD illuminator (Polygon 1000, Mightex, USA). Since the DMD modulates light through binary (on/off) micromirrors, grayscale intensity patterns were generated by spatial dithering over 9×99 \times9 mirror supercells, allowing the traveling wave’s continuous intensity profile to be approximated at the resolution of one supercell. A DAPI filter cube suppressed reflected UV light to prevent camera saturation. Image sequences (2160 ×\times 2560 pixels) were acquired with an sCMOS camera (Andor Zyla) at 10 s−1^{-1}.

Notation

Position in the lab frame is denoted by xx. The wave-comoving phase is

ζ≡mod⁡(x−xwave(t)ℓ+14,1)∈[0,1),(4)\zeta\equiv\operatorname{mod}\left(\frac{x-x_{\mathrm{wave}}(t)}{\ell}+\frac{1}{4},1\right)\in[0,1), \tag*{(4)}

where xwave(t)x_{\mathrm{wave}}(t) is the position of the traveling light pattern and ℓ\ell the wavelength; the offset 1/4 is chosen so that the wave intensity peaks at ζ=0.5\zeta= 0.5 rather than at ζ=0.25\zeta= 0.25, its location in the raw lab-frame coordinate. In this convention the driving profile is

u(ζ)=12(1−cos⁡2πζ),(5)u(\zeta)=\frac{1}{2}(1-\cos2\pi\zeta), \tag*{(5)}

zero at ζ=0\zeta= 0 and ζ=1\zeta= 1, peaking at ζ=0.5\zeta= 0.5. These conventions are used throughout both this Supplemental Material and the main text.

Particle tracking and analysis

Particle positions were detected in each frame and linked into trajectories using custom MATLAB routines. The self-propulsion speed v0v_0 was determined by fitting the particle ensemble’s MSD, computed from 20 s of full-field UV illumination prior to pattern onset, to the ABP model

MSD⁡(Δt)=(4D+2v02Dr−1)Δt+2v02Dr−2(e−ΔtDr−1),(6)\operatorname{MSD}(\Delta t) = \left(4D + 2v_0^2D_r^{-1}\right)\Delta t + 2v_0^2D_r^{-2}\left(e^{-\Delta tD_r}-1\right), \tag*{(6)}

with DD and DrD_r as free parameters, giving D=0.11 μm2/sD = 0.11\,\mathrm{\mu m^2/s} and Dr=1/τr=0.16 s−1D_r = 1/\tau_r = 0.16\,\mathrm{s^{-1}} (τr=6.3 s\tau_r = 6.3\,\mathrm{s}), consistent with the values reported in Ref. [15].

Instantaneous velocities were extracted as a function of wave phase using a fixed-lag, phase-binned estimator. For each particle track, the velocity was computed over a fixed window of nstep=3n_{\mathrm{step}} = 3 frames rather than frame-to-frame, short enough that the particle does not appreciably sweep across the wave during the measurement, but long enough to average out pixel-level tracking noise. Each sample was tagged with the wave-comoving phase ζ\zeta, evaluated at the window’s starting position and time using the logged position of the light pattern. All (ζ,vx,vy)(\zeta,v_x,v_y) samples from every particle and every valid window across a recording were pooled into a single set, from which the speed ∣v∣=vx2+vy2|v| = \sqrt{v_x^2+v_y^2} was also computed per sample. The pooled samples were then binned by ζ\zeta into 30 equal-width bins spanning one full period, and the mean and standard error within each bin yield the phase-resolved velocity profiles vζv_\zeta, vy(ζ)v_y(\zeta), and ∣v∣(ζ)|v|(\zeta) reported throughout the main text. No drift correction was applied to these phase-resolved profiles.

Net drift velocities are only a few hundred nm s−1\mathrm{nm\,s^{-1}}, so even a small background motion of the sample or stage can bias them. We therefore subtract this background motion before computing the net drift velocity. Because it need not be uniform across the field of view, we estimate it locally. We divide the field of view into a 2×22 \times2 grid of boxes, with neighboring boxes overlapping by 50%. Within each box, we collect the velocities of all particles whose displacement starts inside it, computed over lag times from 1 s to 11 s and pooled together. The velocity distribution in each box is dominated by a peak from particles that are not actively transported, and this peak is offset from zero only by the background motion. We locate it as the maximum of a Gaussian kernel density estimate closest to zero velocity, and take its position as the box’s background velocity. Subtracting this background velocity from all velocities in the box and averaging gives the box’s corrected net drift velocity. The net drift velocity and its uncertainty are the mean and standard deviation of this quantity across all boxes.

Simulation model

Particle dynamics were modeled using overdamped active Brownian particles (ABPs) in two dimensions. Each particle ii (i=1,…,Ni = 1,\ldots,N) is described by its position ri\mathbf{r}_i and orientation angle θi\theta_i, evolving according to

r˙i=va,i(t)p^i+2D ηi(t),(7)\dot{\mathbf{r}}_i = v_{a,i}(t)\hat{\mathbf{p}}_i + \sqrt{2D}\,\eta_i^{(t)}, \tag*{(7)}
θ˙i=2Dr ηi(r),(8)\dot{\theta}_i = \sqrt{2D_r}\,\eta_i^{(r)}, \tag*{(8)}

where va,iv_{a,i} is the particle’s active (self-propulsion) speed, p^i=(cos⁡θi,sin⁡θi)\hat{\mathbf{p}}_i = (\cos\theta_i,\sin\theta_i) is the propulsion direction, DD and DrD_r are the translational and rotational diffusion coefficients, and the noise terms ηi(t)\eta_i^{(t)}, ηi(r)\eta_i^{(r)} are independent Gaussian white-noise processes with zero mean and unit-time correlator; free rotational diffusion gives each particle a persistence time τr=1/Dr\tau_r = 1/D_r. Periodic boundary conditions were applied in both directions.

The traveling activity wave sets a local target speed

utarget,i(t)=12umax⁡[1+sin⁡(2π(xi−ct)ℓ)],(9)u_{\mathrm{target},i}(t) = \frac{1}{2}u_{\max}\left[1+\sin\left(\frac{2\pi(x_i-ct)}{\ell}\right)\right], \tag*{(9)}

with wave speed cc and wavelength ℓ\ell. Rather than adapting to utargetu_{\mathrm{target}} instantaneously, each particle’s active speed relaxes toward it with a first-order lag,

dva,idt=utarget,i(t)−va,iτi,τi={τonif utarget,i(t)≥va,i(t),τoffif utarget,i(t)<va,i(t),(10)\frac{dv_{a,i}}{dt} = \frac{u_{\mathrm{target},i}(t)-v_{a,i}}{\tau_i}, \qquad \tau_i = \begin{cases} \tau_{\mathrm{on}} & \text{if } u_{\mathrm{target},i}(t) \ge v_{a,i}(t),\\ \tau_{\mathrm{off}} & \text{if } u_{\mathrm{target},i}(t) < v_{a,i}(t), \end{cases} \tag*{(10)}

integrated exactly over each time step. All simulations reported here used τon=0\tau_{\mathrm{on}} = 0 (instantaneous rise), with τoff≡τ\tau_{\mathrm{off}} \equiv\tau the free delay parameter carrying the entire asymmetric response. The maximum active speed umax⁡u_{\max} (0 μm s−1\mu\mathrm{m}\,\mathrm{s}^{-1} to 10 μm s−1\mu\mathrm{m}\,\mathrm{s}^{-1}), wave speed cc (0 μm s−1\mu\mathrm{m}\,\mathrm{s}^{-1} to 20 μm s−1\mu\mathrm{m}\,\mathrm{s}^{-1}), and delay τ\tau (0 s to 8 s) were swept at fixed wavelength ℓ=200 μm\ell= 200\,\mu\mathrm{m}. Simulations were performed in a periodic square domain of side L=800 μmL = 800\,\mu\mathrm{m} containing N=640N = 640 particles (number density N/L2=0.001 μm−2N/L^2 = 0.001\,\mu\mathrm{m}^{-2}). Translational and rotational diffusion coefficients were fixed via the Stokes–Einstein relation for a particle radius r=1.03 μmr = 1.03\,\mu\mathrm{m}: D=0.11 μm2/sD = 0.11\,\mu\mathrm{m}^2/\mathrm{s} and Dr=0.16 s−1D_r = 0.16\,\mathrm{s}^{-1} (persistence time τr=1/Dr≈6.25\tau_r = 1/D_r \approx6.25 s), using kBT=0.004 09k_{\mathrm{B}}T = 0.004\,09 pN μm\mu\mathrm{m} and mobility μ=26.9 μm pN−1s−1\mu= 26.9\,\mu\mathrm{m}\,\mathrm{pN}^{-1}\mathrm{s}^{-1}. Equations of motion were integrated using the Euler–Maruyama scheme with a time step Δt=1\Delta t = 1 ms, chosen to satisfy Δt≪τr\Delta t \ll\tau_r and Δt≪ℓ/c\Delta t \ll\ell/c. Each simulation ran for 10610^6 steps (total simulated time T=1000T = 1000 s), with configurations saved every 100 steps (10410^4 saved frames). Simulated trajectories were then converted to the wave-comoving phase ζ\zeta and phase-binned identically to the experimental tracking data (see Particle tracking and analysis, above), to extract the local mean velocity vζv_\zeta, density, and polarization.

Propulsion response to illumination switching

The active speed’s response to a step change in UV illumination was characterized separately from the traveling-wave experiments, under spatially uniform illumination. Switching illumination on from a dark baseline, the active speed rises to its steady-state value within 0.4 s, close to the 0.3 s window used for velocity extraction, so this rise cannot be resolved as a genuine exponential relaxation and is only bounded from above by this timescale. This transient was normalized and averaged over three illumination cycles across four cells, with shaded regions in Figure S1 indicating 95% confidence intervals.

Driving profile measurement

The target-speed profile imposed by the traveling wave is not assumed to be sinusoidal, and is instead measured from the particles’ own phase-resolved speed. As the wave passes over a particle, the particle first experiences increasing intensity up to the wave peak, and then decreasing intensity after it. We measure the driving profile only from the first, rising part, using the window ζ∈[0.45,0.95]\zeta\in[0.45,0.95]. On this half, where the active speed rises with the increasing intensity, the effect of the delay is minimized (Propulsion response to illumination switching, above), so the rising-half speed closely approximates utarget(ζ)u_{\mathrm{target}}(\zeta). On the falling half, in contrast, the speed lags the target by the delay τ\tau and therefore reflects both the driving shape and the delay. A further advantage of extracting the driving profile from the particles’ speed rather than from the light field is that it automatically captures the profile the particles actually respond to, including any distortion from the DMD dithering and any nonlinearity in the speed–intensity response.

Correction for Rice-distributed speed measurements. The measured speed ∣v∣=hypot⁡(vxmeas,vymeas)|v| = \operatorname{hypot}(v_x^{\mathrm{meas}}, v_y^{\mathrm{meas}}) is reconstructed from finite-difference position increments, with each Cartesian velocity component subject to independent, approximately Gaussian measurement noise of standard deviation σ=2D/Δtlag\sigma= \sqrt{2D/\Delta t_{\mathrm{lag}}}, where DD is the translational diffusion coefficient and Δtlag\Delta t_{\mathrm{lag}} the lag time used for the finite difference. If vxmeas∼N(νcos⁡θ,σ2)v_x^{\mathrm{meas}} \sim\mathcal{N}(\nu\cos\theta,\sigma^2) and vymeas∼N(νsin⁡θ,σ2)v_y^{\mathrm{meas}} \sim\mathcal{N}(\nu\sin\theta,\sigma^2) for a true active speed ν\nu, the magnitude ∣v∣|v| follows a Rice distribution Rice⁡(ν,σ)\operatorname{Rice}(\nu,\sigma), whose mean is systematically biased above ν\nu:

E[∣v∣]=σπ2e−s[(1+2s)I0(s)+2sI1(s)],s≡ν24σ2,(11)\mathbb{E}[|v|] = \sigma\sqrt{\frac{\pi}{2}} e^{-s} \left[(1+2s)I_0(s) + 2sI_1(s)\right], \qquad s \equiv\frac{\nu^2}{4\sigma^2}, \tag*{(11)}

with I0,I1I_0,I_1 modified Bessel functions of the first kind.

Shared driving profile. We pool the rising-half, phase-binned mean speed from 11 recordings spanning c=4.1 μm s−1c = 4.1\ \mu\mathrm{m}\,\mathrm{s}^{-1} to 7.7 μm s−17.7\ \mu\mathrm{m}\,\mathrm{s}^{-1} and v0=4.29 μm s−1v_0 = 4.29\ \mu\mathrm{m}\,\mathrm{s}^{-1} to 7.59 μm s−17.59\ \mu\mathrm{m}\,\mathrm{s}^{-1}, each normalized by its own v0v_0, the active speed obtained from the MSD fit under full-field illumination before pattern onset (Particle tracking and analysis, above). We describe all recordings with one shared dimensionless shape f(ζ)f(\zeta), so that utarget(ζ)=v0f(ζ)u_{\mathrm{target}}(\zeta) = v_0 f(\zeta), and represent ff by a second-order Fourier series. Its coefficients are determined by a least-squares fit of the forward-biased model speed, i.e. v0f(ζ)v_0f(\zeta) passed through Eq. (11), to the measured mean speed in each phase bin. The resulting driving profile, Figure S2, is used for all theory–experiment comparisons (Figure S4).

Figure S2. Driving profile from the rising half. Gray: mean speed ∣v∣/v0|v|/v_0 over the fit window ζ∈[0.45,0.95]\zeta\in[0.45, 0.95], for every recording used in the fit, each normalized by its own pre-pattern active speed v0v_0. Purple: the shared driving profile v0f(ζ)v_0f(\zeta), first forward-biased through Eq. (11) with each recording’s noise level and then normalized by v0v_0, one line per recording. The collapse of recordings with different cc and v0v_0 onto one curve supports a single driving shape whose amplitude is set by v0v_0 alone.

Model

Moment expansion. The particle’s Langevin dynamics in the wave-comoving frame z≡x−ctz \equiv x - ct, and the resulting Fokker–Planck equation for the joint probability density ψ(z,θ,v,t)\psi(z,\theta,v,t),

∂ψ∂t=−∂∂z[(vcos⁡θ−c)ψ]+D∂2ψ∂z2+Dr∂2ψ∂θ2+∂∂v[(v−u(z)τ(z,v))ψ],(12)\frac{\partial\psi}{\partial t} = -\frac{\partial}{\partial z}\left[(v\cos\theta-c)\psi\right] + D\frac{\partial^2\psi}{\partial z^2} + D_r\frac{\partial^2\psi}{\partial\theta^2} + \frac{\partial}{\partial v}\left[\left(\frac{v-u(z)}{\tau(z,v)}\right)\psi\right], \tag*{(12)}

are given in the main text. We reduce Eq. (12) by angular-moment truncation to a small, closed linear system, solved directly as a single sparse linear algebra problem in the wave-comoving frame. ψ\psi is expanded in harmonics of θ\theta. Since the wave profile is translation-invariant along the yy axis, ψ\psi contains no sin⁡(kθ)\sin(k\theta) terms, and the expansion is a pure cosine series,

ψ(z,θ,v,t)=12π[Ψ0(z,v,t)+2∑k=1∞Ψk(z,v,t)cos⁡(kθ)],(13)\psi(z,\theta,v,t)=\frac{1}{2\pi}\left[\Psi_0(z,v,t)+2\sum_{k=1}^{\infty}\Psi_k(z,v,t)\cos(k\theta)\right], \tag*{(13)}

with coefficients

Ψ0(z,v)=∫ψ dθ,Ψ1(z,v)=∫cos⁡θ ψ dθ,Ψ2(z,v)=∫cos⁡2θ ψ dθ,(14)\Psi_0(z,v)=\int\psi\,\mathrm{d}\theta,\qquad\Psi_1(z,v)=\int\cos\theta\,\psi\,\mathrm{d}\theta,\qquad\Psi_2(z,v)=\int\cos2\theta\,\psi\,\mathrm{d}\theta, \tag*{(14)}

the density, and the polar and nematic order parameters of the heading distribution, respectively: Ψ1\Psi_1 is nonzero under a genuine head-tail asymmetry, with more particles oriented one way along an axis than in the opposite direction, while Ψ2\Psi_2, familiar from liquid-crystal order, is nonzero under alignment onto an axis without distinguishing its two directions, and can be nonzero even when Ψ1=0\Psi_1=0. No explicit alignment interaction appears in Eq. (12); persistent orientation enters only through advection and isotropic rotational diffusion. We look for stationary states in the wave-comoving frame, so we set ∂t∣z→0\partial_t|_z\to0 throughout. Because {1,cos⁡θ,cos⁡2θ,…}\{1,\cos\theta,\cos2\theta,\ldots\} are mutually orthogonal on [0,2π)[0,2\pi), projecting Eq. (12) onto cos⁡(kθ)\cos(k\theta) recovers the equation obeyed by Ψk\Psi_k directly from the series above, rather than a separate calculation: multiplying Eq. (12) by cos⁡(kθ)\cos(k\theta) for k=0,1,2k=0,1,2 and integrating over θ\theta, using cos⁡θcos⁡(kθ)=12[cos⁡((k−1)θ)+cos⁡((k+1)θ)]\cos\theta\cos(k\theta)=\frac{1}{2}[\cos((k-1)\theta)+\cos((k+1)\theta)] and ∫cos⁡(kθ) ∂θ2ψ dθ=−k2∫cos⁡(kθ) ψ dθ\int\cos(k\theta)\,\partial_\theta^2\psi\,\mathrm{d}\theta=-k^2\int\cos(k\theta)\,\psi\,\mathrm{d}\theta, and closing the hierarchy by setting Ψ3≡0\Psi_3\equiv0, gives

−∂z[vΨ1−cΨ0]+D∂z2Ψ0+∂v[v−u(z)τ(z,v)Ψ0]=0,(15)-\partial_z\left[v\Psi_1-c\Psi_0\right]+D\partial_z^2\Psi_0+\partial_v\left[\frac{v-u(z)}{\tau(z,v)}\Psi_0\right]=0, \tag*{(15)}
−∂z[12v(Ψ0+Ψ2)−cΨ1]+D∂z2Ψ1−DrΨ1+∂v[v−u(z)τ(z,v)Ψ1]=0,(16)-\partial_z\left[\frac{1}{2}v(\Psi_0+\Psi_2)-c\Psi_1\right]+D\partial_z^2\Psi_1-D_r\Psi_1+\partial_v\left[\frac{v-u(z)}{\tau(z,v)}\Psi_1\right]=0, \tag*{(16)}
−∂z[12vΨ1−cΨ2]+D∂z2Ψ2−4DrΨ2+∂v[v−u(z)τ(z,v)Ψ2]=0.(17)-\partial_z\left[\frac{1}{2}v\Psi_1-c\Psi_2\right]+D\partial_z^2\Psi_2-4D_r\Psi_2+\partial_v\left[\frac{v-u(z)}{\tau(z,v)}\Psi_2\right]=0. \tag*{(17)}

Why the nematic order matters. Each equation is sourced by a spatial derivative of the moment one order below: Ψ1\Psi_1’s equation is driven by ∂z[vΨ0]\partial_z[v\Psi_0], and Ψ2\Psi_2’s in turn by ∂z[vΨ1]\partial_z[v\Psi_1], each of order umax⁡Ψ0/ℓu_{\max}\Psi_0/\ell and umax⁡Ψ1/ℓu_{\max}\Psi_1/\ell respectively, since ψ\psi varies over the wavelength ℓ\ell. Balancing this forcing against the dominant relaxation term at each order, DrΨ1D_r\Psi_1 and 4DrΨ24D_r\Psi_2, gives the rough scaling Ψ1/Ψ0∼Ψ2/Ψ1∼umax⁡/(ℓDr)=ℓp/ℓ\Psi_1/\Psi_0\sim\Psi_2/\Psi_1\sim u_{\max}/(\ell D_r)=\ell_p/\ell, so Ψ2/Ψ0∼(ℓp/ℓ)2\Psi_2/\Psi_0\sim(\ell_p/\ell)^2, with ℓp≡umax⁡/Dr\ell_p\equiv u_{\max}/D_r the persistence length. The cascade is therefore weak, and truncating at low order safe, when ℓp≪ℓ\ell_p\ll\ell, but strong once ℓp≳ℓ\ell_p\gtrsim\ell, matching the same intermediate-scale regime in which higher-order closures become necessary for plain active Brownian particles [34]. This delay-aware, steady-state (z,v)(z,v) system, closed at Ψ3≡0\Psi_3\equiv0, is the model used throughout. Closing instead at Ψ2≡0\Psi_2\equiv0 predicts the nematic order to be exactly zero, independent of any residual-based accuracy comparison; we test this directly against real simulated trajectories in Figure S3, which indeed shows the RMS nematic order growing with increasing ℓp/ℓ\ell_p/\ell (colorbar), and confirms that closing the hierarchy at Ψ3≡0\Psi_3\equiv0 reproduces this simulated RMS nematic order to a good agreement. We map the predicted size of the effect directly across the (ℓp,ℓ)(\ell_p,\ell) plane in Figure S3, where the RMS nematic order is largest for large ℓp\ell_p and small ℓ\ell, consistent with the scaling Ψ2/Ψ0∼(ℓp/ℓ)2\Psi_2/\Psi_0\sim(\ell_p/\ell)^2 derived above.

Zero-delay limit. In this limit, the active speed is adiabatically eliminated to the local target, v=u(z)v=u(z), and vv is no longer an independent coordinate: substituting it directly into Eq. (12) removes the ∂v[…]\partial_v[\ldots] term entirely, and repeating the same angular-moment projection on this simplified equation gives a closed system of ordinary differential equations in zz alone, obtained from Eqs. (15)–(17) by dropping the delay term and replacing vv with u(z)u(z) throughout.

Delay-aware case. When τ>0\tau>0, the active speed can no longer be eliminated as v=u(z)v=u(z): it remains a genuine internal coordinate, and Eqs. (15)–(17) retain their full vv-dependence through τ(z,v)\tau(z,v), which switches between τon\tau_{\mathrm{on}} and τoff\tau_{\mathrm{off}} depending on whether the local target speed is rising or falling relative to vv. This makes the delay-aware steady-state system a genuine two-dimensional problem in (z,v)(z,v), not reducible back to the one-dimensional system above.

Discretization. Both cases are linear, since the moments couple only linearly, so after discretization the entire problem reduces to a sparse linear system Au=b\mathbf{A}\mathbf{u}=\mathbf{b}. The zz-derivatives (diffusion and advection) are approximated with central differences on a periodic grid over one wavelength, z∈[0,ℓ)z\in[0,\ell), in both cases. The additional vv-advection term in the delay-aware case is discretized instead with an upwind finite-volume scheme, since that direction is not periodic and has physical boundaries at v=0v=0 and v=umax⁡v=u_{\max}; central differencing is unstable for a purely advective term, whereas upwinding (drawing the flux from whichever neighboring cell the drift originates from) is stable. In both cases, resulting matrix AA is singular by construction, since the moment equations fix only the shape of the distribution and not its overall scale; one row of AA is therefore overwritten by the normalization constraint (∫0ℓΨ0 dz=1\left(\int_0^\ell\Psi_0\,\mathrm{d}z=1\right. in the zero-delay case, ∫0ℓ∫0umax⁡Ψ0 dv dz=1)\left.\int_0^\ell\int_0^{u_{\max}}\Psi_0\,\mathrm{d}v\,\mathrm{d}z=1\right) in the delay-aware case), with the corresponding entry of b\mathbf{b} set to 1. The system is then solved directly by sparse LU decomposition, with no iteration or time-stepping required.

Observables. Once Ψ0\Psi_0, Ψ1\Psi_1, Ψ2\Psi_2 are known, all reported quantities follow by quadrature over zz: the density Ψ0(z)\Psi_0(z), the polar order Ψ1(z)\Psi_1(z), the flux q(z)=∫vΨ1 dvq(z)=\int v\Psi_1\,\mathrm{d}v (or q(z)=v(z)Ψ1(z)q(z)=v(z)\Psi_1(z) in the zero-delay case, where no separate vv-integral is needed), and the local mean velocity vz(z)=q(z)/Ψ0(z)v_z(z)=q(z)/\Psi_0(z). For comparison with experiment and simulation we report these on the dimensionless, periodic phase ζ≡z/ℓ\zeta\equiv z/\ell (Notation, above) used throughout the main text. The density picks up a Jacobian factor to remain normalized, ρ(ζ)≡ℓΨ0(z=ℓζ)\rho(\zeta)\equiv\ell\Psi_0(z=\ell\zeta), so that ∫01ρ(ζ) dζ=1\int_0^1\rho(\zeta)\,\mathrm{d}\zeta=1; the mean velocity, an intensive ratio, is unaffected by this rescaling, vζ≡vz(z=ℓζ)v_\zeta\equiv v_z(z=\ell\zeta). The net drift velocity then follows either as vtransport=∫01ρ(ζ)vζ dζv_{\mathrm{transport}}=\int_0^1\rho(\zeta)v_\zeta\,\mathrm{d}\zeta, matching the main text since ρvζ=q(ℓζ)\rho v_\zeta=q(\ell\zeta), or equivalently, without first forming ρ(ζ)\rho(\zeta), directly as vtransport=∫0ℓq(z) dzv_{\mathrm{transport}}=\int_0^\ell q(z)\,\mathrm{d}z. Because the system is solved exactly (given DD, DrD_r, cc, ℓ\ell, umax⁡u_{\max}, and τ\tau) rather than sampled stochastically, this constitutes a closed-form numerical solution of the truncated-moment system rather than a simulation.

Model validation

Each run is reduced to the peak and trough of its vζ(ζ)v_\zeta(\zeta) curve, and compared against the theory evaluated for that same run (Figure S4).

References

  1. [5]R. E. Goldstein, Physical Review Letters 77, 775–778 (1996).
  2. [6]C. Maggi, L. Angelani, G. Frangipane, and R. Di Leonardo, Soft Matter 14, 4958–4962 (2018).
  3. [7]A. Geiseler, P. Hänggi, F. Marchesoni, C. Mulhern, and S. Savel’ev, Physical Review E 94, 10.1103/physreve.94.012613 (2016).
  4. [8]A. Geiseler, P. Hänggi, and F. Marchesoni, Entropy 19, 97 (2017).
  5. [9]A. Geiseler, P. Hänggi, and F. Marchesoni, Scientific Reports 7, 10.1038/srep41884 (2017).
  6. [10]C. Lozano and C. Bechinger, Nature Communications 10, 10.1038/s41467-019-10535-z (2019).
  7. [11]J. Palacci, S. Sacanna, S.-H. Kim, G.-R. Yi, D. J. Pine, and P. M. Chaikin, Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 372, 20130372 (2014).
  8. [12]M. Zehavi, D. Sofer, T. Miloh, O. D. Velev, and G. Yossifon, Physical Review Applied 18, 10.1103/physrevapplied.18.024060 (2022).
  9. [13]G. Frangipane, D. Dell’Arciprete, S. Petracchini, C. Maggi, F. Saglimbeni, S. Bianchi, G. Vizsnyiczai, M. L. Bernardini, and R. Di Leonardo, eLife 7, 10.7554/elife.36608 (2018).
  10. [14]J. Arlt, V. A. Martinez, A. Dawson, T. Pilizota, and W. C. K. Poon, Nature Communications 9, 10.1038/s41467-018-03161-8 (2018).
  11. [15]U. Töpfer, M. R. Bailey, S. Schreiber, F. Paratore, and L. Isa, ACS Nano 19, 39210–39219 (2025).
  12. [16]S. A. M. Loos and S. H. L. Klapp, Journal of Statistical Physics 177, 95–118 (2019).
  13. [17]M. Feito and F. Cao, Physica A: Statistical Mechanics and its Applications 387, 4553–4559 (2008).
  14. [18]S. A. M. Loos, R. Gernert, and S. H. L. Klapp, Physical Review E 89, 10.1103/physreve.89.052136 (2014).
  15. [19]M. Mijalkov, A. McDaniel, J. Wehr, and G. Volpe, Physical Review X 6, 10.1103/physrevx.6.011008 (2016).
  16. [20]M. Leyman, F. Ogemark, J. Wehr, and G. Volpe, Physical Review E 98, 10.1103/physreve.98.052606 (2018).
  17. [21]M. A. Fernandez-Rodriguez, F. Grillo, L. Alvarez, M. Rathlef, I. Buttinoni, G. Volpe, and L. Isa, Nature Communications 11, 10.1038/s41467-020-17864-4 (2020).
  18. [22]H. Massana-Cid, C. Maggi, G. Frangipane, and R. Di Leonardo, Nature Communications 13, 10.1038/s41467-022-30201-1 (2022).
  19. [23]V. Holubec, A. Fischer, G. Volpe, and F. Cichos, Physical Review Research 8, 10.1103/xn9x-ppjx (2026).
  20. [24]V. Holubec and F. Cichos, Feedback control and delayed interactions in active matter (2026).arxiv.org/abs/2603.17894
  21. [25]C. Rein, K. Kroy, and V. Holubec, Reaction delays boost, rectify, and reverse motility in activity landscapes (2026).arxiv.org/abs/2608.03906
  22. [26]A. Boymgreen, J. Schiffbauer, B. Khusid, and G. Yossifon, Current Opinion in Colloid & Interface Science 60, 101603 (2022).
  23. [27]T. M. Squires and M. Z. Bazant, Journal of Fluid Mechanics 560, 65–101 (2006).
  24. [28]J. Palacci, S. Sacanna, A. P. Steinberg, D. J. Pine, and P. M. Chaikin, Science 339, 936–940 (2013).
  25. [29]C. Chen, F. Mou, L. Xu, S. Wang, J. Guan, Z. Feng, Q. Wang, L. Kong, W. Li, J. Wang, and Q. Zhang, Advanced Materials 29, 10.1002/adma.201603374 (2016).
  26. [30]A. Aubret, M. Youssef, S. Sacanna, and J. Palacci, Nature Physics 14, 1114–1118 (2018).
  27. [31]A. Litke, E. J. M. Hensen, and J. P. Hofmann, The Journal of Physical Chemistry C 121, 10153–10162 (2017).
  28. [32]A. Litke, Y. Su, I. Tranca, T. Weber, E. J. M. Hensen, and J. P. Hofmann, The Journal of Physical Chemistry C 121, 7514–7524 (2017).
  29. [33](2026), see Supplemental Material at [URL] for experimental details, derivations, and supporting figures.
  30. [34]T. Gautry, M. Deforet, and P. Illien, Journal of Statistical Mechanics: Theory and Experiment 2026, 013207 (2026).
  31. [35]M. J. Schnitzer, Physical Review E 48, 2553–2568 (1993).
  32. [36]J. Tailleur and M. E. Cates, Physical Review Letters 100, 10.1103/physrevlett.100.218103 (2008).
  33. [37]J. Luczka, R. Burschka, and P. Hänggi, Europhysics Letters (EPL) 31, 431–436 (1995).

Paper details

vlnc:W_3544shgp
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 6

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Anthropic Claude Opus 5.5

View statement
Looking for related papers…
Contents