betaFeedback

WorksPhysics and Astronomy

Inferring Surface Slip in Active Colloids from Flow Fields Using Physics-Informed Neural Networks

Abstract

The directed motion of active colloids is governed by spatial variations in surface chemistry and interfacial stress, yet these properties remain extremely difficult to measure directly. We introduce a physics-informed neural network framework that infers the slip distribution driving propulsion from partial observations of the surrounding flow. By combining sparse fluid velocity measurements with the Stokes equations and boundary constraints, the method reconstructs both the near-surface slip and the full velocity and pressure fields. Validation against analytical solutions and Boundary Element Method calculations for canonical active colloid models shows quantitative agreement in both unbounded and confined geometries. Crucially, the framework recovers the surface slip even when no flow data are available near the particle, demonstrating that accessible bulk measurements encode the interfacial stresses responsible for active motion. These results establish physics-informed inference as a powerful tool for characterizing and ultimately controlling interfacially driven transport in colloidal active matter.

Introduction

Interfacial-driven flows arise from spatial variations in mechanical stress generated by gradients in interfacial tension, temperature, concentration, surface charge, or chemical activity [1, 2, 3, 4]. Because these stresses mediate momentum transfer between interfaces and the surrounding fluid, they govern how soft materials and complex liquids move, reorganize, and respond to their environments [5, 6, 7, 8]. A prominent example, and the focus of this work, is colloidal active matter, in which interfacial stresses drive particle self-propulsion. Spatial variations in surface chemistry or interfacial tension generate surface flows that propel particles through viscous fluids [9, 10, 11, 12, 13]. These stresses govern propulsion and shape particle–boundary interactions [14, 15, 16, 17, 18, 19, 20, 21, 22, 23], particle–particle dynamics [24, 25, 26, 27, 28, 29, 30], and responses to external fields [31, 32, 33, 34, 35, 36, 37]. More broadly, resolving near-surface flow and slip is essential for understanding the coupling between mechanical transport and biochemical activity, as seen in cortical flows of biological systems such as Drosophila embryos, where spatially organized surface contractility coordinates nuclear positioning and cell-cycle oscillations [38, 39, 40]. Quantitatively resolving these surface properties is therefore essential for predicting transport and nonequilibrium organization in active matter.

Despite their importance, quantities such as slip velocity, phoretic mobility, and local interfacial stress remain difficult to measure. Bulk flow fields can be resolved with modern optical imaging techniques [41, 42], but measurements near surfaces are often noisy or incomplete because of optical distortions and tracer–boundary interactions [43, 44]. Direct stress measurements are even more limited, since traction force microscopy and interfacial rheometry provide only substrate-restricted or spatially averaged information [45, 46]. As a result, the interfacial properties that govern propulsion often remain experimentally inaccessible.

Inferring these hidden boundary conditions from surrounding flow data requires solving an inverse Stokes problem that is typically ill-posed, highly sensitive to noise, and computationally demanding [47]. Classical numerical approaches, such as finite and boundary element methods (FEM/BEM), excel at forward simulations, where boundary conditions are prescribed, but are not designed to infer them from partial data. These limitations motivate frameworks that combine physical constraints with available measurements to extract interfacial information directly from flow fields.

Physics-informed neural networks (PINNs) provide one such route. By embedding the governing equations into the training loss together with experimental or simulated data, PINNs enforce physical consistency even when measurements are sparse or incomplete [48, 49, 50, 51]. They have been used to reconstruct flow fields from incomplete observations [52, 53, 54, 55], infer pressure from MRI data [56], and identify constitutive laws in complex materials [57]. These prior studies illustrate how physics-constrained learning can be a powerful tool for solving inverse problems that challenge traditional numerical approaches.

Here, we introduce a PINN-based framework that infers the surface slip velocity of an active colloid directly from sparse measurements of the surrounding flow (Figure 1). The network reconstructs the full velocity and pressure fields by enforcing the Stokes equations and boundary conditions, producing solutions consistent with both the data and the underlying hydrodynamics. This mesh-free approach is particularly suited to situations where near-surface flow cannot be measured accurately. By inferring the hidden slip distribution and its associated flow, the method quantitatively links experimentally accessible hydrodynamic fields to the interfacial mechanisms that generate propulsion.

Flow around an active colloid and physics-informed neural network framework

Figure 1. (a) Bulk flow around an active colloid can be measured accurately, but data within the near-surface region (red hatched band) are typically noisy or missing, obscuring the slip velocity and interfacial stresses that generate propulsion. (b) Spatial coordinates r=(x,y,z)\boldsymbol{r}=(x,y,z) serve as inputs, and the network outputs (uxnn,uynn,uznn,pnn)(u_x^{\mathrm{nn}},u_y^{\mathrm{nn}},u_z^{\mathrm{nn}},p^{\mathrm{nn}}). Physics-based losses enforce the Stokes equations, incompressibility, force- and torque-free conditions, boundary constraints, and no-slip walls when present, while a data-loss term penalizes deviations from measured velocities. Minimizing the total loss yields physically consistent reconstructions of the flow field and surface slip distribution.

The remainder of the manuscript is organized as follows. We outline the hydrodynamics of active colloids and introduce the squirmer and phoretic Janus particle models. We then describe the PINN formulation and validate it using analytical solutions and BEM simulations. Finally, we discuss the broader implications of this framework and its extensions to more complex systems.

**Active Colloid Hydrodynamics.—**To model the dynamics of an active colloid, we consider a particle of radius aa moving in a low Reynolds number fluid described by the Stokes equations

η∇2u−∇p=0,∇⋅u=0,(1)\eta\nabla^2\boldsymbol{u}-\nabla p=0,\qquad\nabla\cdot\boldsymbol{u}=0, \tag*{(1)}

where η\eta, u\boldsymbol{u}, and pp denote the dynamic viscosity, fluid velocity, and pressure, respectively. For a free particle, the boundary condition at infinity is u→0\boldsymbol{u}\to0, and for confining boundaries, we impose no slip at the wall. At the particle surface, the fluid velocity satisfies

u=U+Ω×(r−r0)+us,(2)\boldsymbol{u}=\boldsymbol{U}+\boldsymbol{\Omega}\times(\boldsymbol{r}-\boldsymbol{r}_0)+\boldsymbol{u}_s, \tag*{(2)}

where U\boldsymbol{U} and Ω\boldsymbol{\Omega} are the translation and angular velocities of the particle, which can be probed experimentally, and us\boldsymbol{u}_s is the slip velocity. The particle satisfies the force- and torque-free conditions

F=∫Sσ⋅n^ dS=0,τ=∫S(r−r0)×(σ⋅n^) dS=0.(3)\boldsymbol{F}=\int_S\boldsymbol{\sigma}\cdot\hat{\boldsymbol{n}}\,dS=0,\qquad\boldsymbol{\tau}=\int_S(\boldsymbol{r}-\boldsymbol{r}_0)\times(\boldsymbol{\sigma}\cdot\hat{\boldsymbol{n}})\,dS=0. \tag*{(3)}

Here, σ=−pI+η(∇u+∇uT)\boldsymbol{\sigma}=-p\boldsymbol{I}+\eta(\nabla\boldsymbol{u}+\nabla\boldsymbol{u}^{\mathsf{T}}) is the hydrodynamic stress tensor, n^\hat{\boldsymbol{n}} the surface normal, and integration is taken over the entire particle surface SS. These constraints, together with the Lorentz reciprocal theorem,

yield the translational and angular velocities, U\boldsymbol{U} and Ω\boldsymbol{\Omega}, for a given slip velocity distribution us\boldsymbol{u}_s [3].

**Squirmer Model.—**A standard model for the slip velocity of an active colloid is the squirmer model, in which

us=B1sin⁡θ+B2sin⁡θcos⁡θ,(4)u_s=B_1\sin\theta+B_2\sin\theta\cos\theta, \tag*{(4)}

where θ\theta is the polar angle relative to the swimming direction d^\hat{\boldsymbol{d}}, and ∣U∣=2∣B1∣/3|U|=2|B_1|/3 is the swimming speed. In this work, we fix B1=1B_1=1. The dimensionless squirmer parameter βs=B2/∣B1∣\beta_s=B_2/|B_1| classifies the flow field generated by the particle as a puller (βs>0\beta_s>0), pusher (βs<0\beta_s<0), or neutral swimmer (βs=0\beta_s=0). The Stokes equations for a squirmer in an unbounded domain admit an analytical solution, which we use as a reference [58, 59]. We also employ BEM to compute flow fields, as it is an accurate, established method for Stokes flows [18, 60].

**Phoretic Janus Particle Model.—**For a phoretic Janus particle, we consider a generic model in which one hemisphere emits solute while the other hemisphere consumes it [25, 61]. This model describes a broad class of chemically driven active colloids, including enzyme-coated particles, thermoresponsive microgels, and bimetallic colloids [9, 62, 63, 64, 65, 66, 67]. The slip velocity resulting from solute gradients along the surface of the particle is

us=−b(n^)∇sC∣s,(5)\boldsymbol{u}_s=-b(\hat{\boldsymbol{n}})\nabla_s C\big|_s, \tag*{(5)}

where CC is the solute concentration, ∇s=(I−n^n^)⋅∇\nabla_s=(\boldsymbol{I}-\hat{\boldsymbol{n}}\hat{\boldsymbol{n}})\cdot\nabla is the surface-gradient operator. The phoretic mobility b(n^)b(\hat{\boldsymbol{n}}) depends on particle-solute interactions and may be positive or negative [68, 69]. We treat b(n^)b(\hat{\boldsymbol{n}}) as constant on each hemisphere, denoted by beb_e and bab_a, and define the mobility ratio βp=ba/be\beta_p=b_a/b_e. We assume a purely diffusive solute so that the solute field obeys the continuity equation ∂tC=D∇2C\partial_t C=D\nabla^2 C, where DD is the solute diffusivity. In our calculations, we assume solute relaxation is fast compared to particle motion, and the concentration field is treated as quasisteady, and satisfies the Laplace equation D∇2(C−C∞)=0D\nabla^2(C-C_\infty)=0, with C∞C_\infty the far-field concentration. We provide additional model details in the End Matter for phoretic Janus particles.

PINN Framework

Figure 1(b) gives an overview of the PINN framework used to infer the near-surface flow field and interfacial slip distribution from partial velocity data. A fully connected neural network with trainable weights and biases (W,b)(W,b) learns a mapping from spatial coordinates r=(x,y,z)\mathbf{r}=(x,y,z) to the hydrodynamic fields (unn,pnn)=(u(r;W,b),p(r;W,b))(\mathbf{u}_{\mathrm{nn}},p_{\mathrm{nn}})=(\mathbf{u}(\mathbf{r};W,b),p(\mathbf{r};W,b)). These weights and biases are optimized by minimizing a composite loss function that couples physical constraints with available data,

L=λ1LPDE+λ2LBC+λ3LFT+λ4Lwall+λ5Ldata(6)\mathcal{L}=\lambda_1\mathcal{L}_{\mathrm{PDE}}+\lambda_2\mathcal{L}_{\mathrm{BC}}+\lambda_3\mathcal{L}_{\mathrm{FT}}+\lambda_4\mathcal{L}_{\mathrm{wall}}+\lambda_5\mathcal{L}_{\mathrm{data}} \tag*{(6)}

where the hyperparameters λi\lambda_i control their relative weights. The loss terms enforce the incompressible Stokes equations, boundary conditions at particle surface, force- and torque-free constraints on the particle, and no-slip conditions on walls. Finally, Ldata\mathcal{L}_{\mathrm{data}} penalizes deviations between predicted and measured velocities at selected points. The network is implemented in PyTorch and trained with the Adam optimizer and an adaptive learning rate until the total loss saturates below a prescribed tolerance. Details of the architecture, loss terms, and hyperparameters are provided in the End Matter.

PINN Results

We first assess the PINN’s ability to recover hydrodynamic fields in an unbounded domain. Figure 2(a,Figure 2c) compares predictions of the flow field for a puller-type squirmer (βs=5)(\beta_s=5) and a phoretic Janus particle (βp=−10)(\beta_p=-10), respectively. The PINN streamlines (blue) closely match BEM solutions (pink), including near-surface regions (r/a<1.2)(r/a<1.2) where no training data were supplied. The PINN also accurately predicts the surface slip velocity: Figure 2(b,Figure 2d) compares the PINN, BEM, and analytical solutions for squirmers with βs={−5,0,5}\beta_s=\{-5,0,5\} [58, 59] and Janus particles with βp={−10,1,10}\beta_p=\{-10,1,10\}, showing excellent agreement in both magnitude and angular dependence of the slip.

Comparison of PINN-predicted and BEM flow fields for squirmer and phoretic Janus particles

Figure 2. Comparison of PINN-predicted (blue) and BEM (pink) flow fields in the yzyz-plane for squirmer and phoretic Janus particles in the laboratory frame. Panels (a,c) show unbounded cases: a puller squirmer with βs=5\beta_s = 5 and a Janus particle with βp=−10\beta_p = -10. Panels (e,g) show corresponding near-wall configurations, with the particle centered at z0/a=2z_0/a = 2; the squirmer is oriented normal to the wall, and the Janus particle is oriented at rot⁡x=π/3\operatorname{rot}_x = \pi/3. Panels (b,f) report the surface slip velocity for squirmers with βs={−5,0,5}\beta_s = \{-5,0,5\}, and panels (d,h) report slip velocities for Janus particles with βp={−10,1,10}\beta_p = \{-10,1,10\}. In all cases, open symbols denote BEM results and filled symbols denote PINN predictions. For the squirmer, solid curves in panels (b,f) show the analytical slip distribution used as reference. Velocities are normalized by the swimming speed ∣U∣\lvert U\rvert.

To quantify the agreement between the PINN predictions and reference solutions, we compute the relative L2L_2 error between the PINN-predicted and analytical (or BEM) velocity fields, EL2=∥uPINN−uref∥2/∥uref∥2E_{L_2}=\lVert\mathbf{u}_{\mathrm{PINN}}-\mathbf{u}_{\mathrm{ref}}\rVert_2/\lVert\mathbf{u}_{\mathrm{ref}}\rVert_2, where ∥⋅∥2\lVert\cdot\rVert_2 denotes the Euclidean norm evaluated over all grid points. Table 1 reports the errors for squirmers and Janus particles across several values of β\beta. In all cases, the PINN achieves small relative errors in both the flow field and the slip velocity, comparable to those obtained from a BEM calculation. For reference, the BEM solution for the squirmer model in an unbounded domain closely matches the analytical solution, with average L2L_2 errors of 0.018 for the flow field and 0.057 for the slip velocity. These results show that the PINN maintains high accuracy across particle types and activity strengths and can recover the near-surface slip even without training data in this region.

To further assess the robustness and generalization capability of the PINN framework, we extend the analysis. to particles near a no-slip rigid boundary. In this setting, the network includes an additional loss term enforcing zero velocity at the wall. The presence of the boundary introduces strong hydrodynamic interactions that substantially alter both the flow field and the particle’s motion, providing a stringent test of the method.

Wall-induced coupling modifies translational and rotational velocities in a manner that depends on propulsion mechanism, distance from the wall, orientation, and the value of β\beta. In addition, for phoretic Janus particles, proximity to the wall perturbs the solute concentration field, thereby changing the slip distribution when the particle approaches the boundary. In contrast, for squirmers, the wall affects only the hydrodynamic flow field, which can alter the translational and angular velocity of the particle, but the slip velocity remains prescribed by Eq. (4).

Despite these complexities, the PINN accurately captures both the qualitative structure and quantitative details of the near-wall hydrodynamics, showing excellent agreement with the BEM solutions. Figure 2(e–Figure 2h) presents representative predictions for a puller squirmer and a phoretic Janus particle positioned above a wall. In both cases, the PINN reconstructs the wall-modified streamlines with high fidelity, including the enhanced circulation and geometric asymmetry generated by the boundary [Fig. 2(e,g)]. The corresponding slip velocities [Fig. 2(f,h)] also agree closely with BEM, with the PINN reproducing the angular dependence across different squirmer modes βs={5,0,−5}\beta_s = \{5, 0, -5\} and phoretic mobility ratios βp={−10,1,10}\beta_p = \{-10, 1, 10\}. Although no analytical solutions exist in the confined geometry, quantitative comparisons with BEM show that the relative L2L_2 errors for both the flow and slip remain comparable to the unbounded-domain values (Table I). This consistency confirms that the PINN continues to infer near-surface slip and the full velocity field with high accuracy, even under strong hydrodynamic confinement.

SquirmerUnboundedNear-wall
βs\beta_sFlow EL2E_{L_2}Slip EL2E_{L_2}Flow EL2E_{L_2}Slip EL2E_{L_2}
-50.0200.0290.0240.075
00.0200.0070.0730.055
50.0200.0310.0320.080
JanusUnboundedNear-wall
βp\beta_pFlow EL2E_{L_2}Slip EL2E_{L_2}Flow EL2E_{L_2}Slip EL2E_{L_2}
-100.0930.1300.0670.092
10.0420.0440.0590.046
100.0530.0460.0550.063

Table I. Relative L2L_2 errors for flow and slip predictions for squirmer and Janus particles in unbounded and near-wall geometries. Unbounded squirmer errors use analytical solutions as a reference for both flow and slip; near-wall flow errors use BEM as a reference, while slip errors remain referenced to the analytical profiles. For Janus particles, all errors are computed relative to BEM solutions.

**Discussion.—**This work introduces a physics-informed neural network framework for resolving the interfacial mechanisms that generate motion in active colloids. By coupling sparse flow-field measurements with the Stokes equations and appropriate boundary conditions, the method reconstructs near-surface slip velocities and the full velocity and pressure fields with high accuracy. Its performance for both squirmers and phoretic Janus particles, in unbounded domains and near confining boundaries, shows that bulk hydrodynamic observations contain sufficient information to infer the slip velocity responsible for propulsion.

The ability to infer boundary conditions directly from partial observations suggests a broader strategy for characterizing soft and active matter [70]. Incorporating additional physical fields, such as solute concentration or temperature, would enable inference of phoretic mobilities, interfacial stresses, and other transport parameters. Extending the framework to interacting colloids or active droplets would allow data-driven reconstruction of collective hydrodynamic interactions [71, 72]. Integrating experimental flow data, where noise and tracer artifacts are unavoidable, represents an important direction, since PINN-based regularization can extract reliable physical information under non-ideal measurement conditions.

Although developed here for active colloids, the approach is general. It provides a practical strategy for inverse problems in complex fluids where boundary conditions, rheological parameters, or interfacial stresses are not directly accessible. Potential applications include microfluidic design [73, 74], parameter and stress inference in non-Newtonian fluids [75, 76], near-interface particle dynamics [16], and the stability and transport behavior of multiphase and biological interfaces [77, 78, 79]. By linking macroscopic flow observations to microscopic interfacial physics, the framework provides a quantitative path forward in systems where traditional computational or experimental tools are limited.

A remaining challenge, shared with many PINN-based approaches, is handling sharp gradients or discontinuities across interfaces. Standard neural networks favor smooth representations, which can reduce accuracy when the true solution has rapid spatial variation. In our case, this manifests as diminished slip-velocity accuracy for a Janus particle very close to a wall (z0/a<1.4z_0/a < 1.4), where the slip profile exhibits steep spatial variations. In contrast, the framework performs well for squirmers at comparable separations (z0/a>1.1z_0/a > 1.1). Recent advances including domain decomposition [80, 81], specialized architectures [82, 83], and interface-aware activation functions [84] offer potential remedies. As these developments mature, they are likely to extend the range and robustness of physics-informed inference for complex interfacial flows.

End Matter

Phoretic Janus Particle: Concentration Field—For the phoretic Janus particle, we consider a particle with Janus balance χ=0\chi= 0, corresponding to a half-coated configuration. Mass conservation requires the emission and absorption fluxes QeQ_e and QaQ_a to satisfy SeQe−SaQa=0S_e Q_e - S_a Q_a = 0, where SeS_e and SaS_a denote the areas of the emitting and absorbing hemispheres. At steady state this implies Qa/Qe=(1+χ)/(1−χ)Q_a/Q_e = (1 + \chi)/(1 - \chi). The solute boundary condition on the particle surface is

−Dn^⋅∇C(r)=H(n^⋅d^)Qe+[1−H(n^⋅d^)]Qa,-D \hat{\mathbf{n}} \cdot\nabla C(\mathbf{r}) = H(\hat{\mathbf{n}} \cdot\hat{\mathbf{d}}) Q_e + [1 - H(\hat{\mathbf{n}} \cdot\hat{\mathbf{d}})] Q_a,

where n^\hat{\mathbf{n}} is the outward normal, d^\hat{\mathbf{d}} is the Janus orientation, and H(x)H(x) is a smooth indicator function. The far-field condition is C(r)→C∞C(\mathbf{r}) \to C_\infty as ∣r∣→∞|\mathbf{r}| \to\infty; in wall-bounded domains, CC satisfies a no-flux boundary condition at the wall. To avoid discontinuities at the emitting–absorbing interface, we smooth the phoretic mobility profile as

b(n^)be=H(n^⋅d^)+[1−H(n^⋅d^)]βp,\frac{b(\hat{\mathbf{n}})}{b_e} = H(\hat{\mathbf{n}} \cdot\hat{\mathbf{d}}) + [1 - H(\hat{\mathbf{n}} \cdot\hat{\mathbf{d}})]\beta_p,

with beb_e the mobility on the emitting side and βp=ba/be\beta_p = b_a/b_e the mobility ratio. We adopt a sigmoid transition,

H(x)=12[1+tanh⁡(3(x−χ))],H(x) = \frac{1}{2}[1 + \tanh(3(x - \chi))],

to regularize the hemispherical discontinuity. In unbounded geometries, analytical solutions of the Laplace equation for CC are available [18, 37]. For the general wall-bounded configurations considered here, we compute the solute field using BEM.

BEM Calculation Details—Boundary element method (BEM) simulations are used to compute the solute concentration, fluid velocity fields, and particle kinematics. We solve the Laplace and Stokes equations in three dimensions using their boundary-integral formulations, implemented with BEMLIB routines in Fortran [60] and extended to handle active Janus particles [17, 25]. These calculations provide the reference solutions used for validation in both unbounded and wall-bounded geometries.

PINN Implementation Details—The Physics-Informed Neural Network (PINN) framework was implemented in PyTorch 2.2.1. Separate neural networks were assigned to each output variable to improve convergence and stability. For rigid particles, the computational domain was divided into two regions—the particle surface and the exterior fluid—each represented by independent networks predicting the velocity components and pressure, yielding a total of eight networks. This modular design ensures continuity across interfaces and facilitates extensions to more complex geometries and boundary conditions. All spatial derivatives required to evaluate the Stokes residuals and stress tensors were computed via automatic differentiation.

The network architecture consisted of nine fully connected hidden layers with 128 neurons per layer. Weights were initialized using PyTorch’s default Kaiming-uniform initialization, and the Swish activation function f(x)=xsigmoid⁡(x)f(x) = x\operatorname{sigmoid}(x) was used throughout to ensure smooth second derivatives required by the governing equations. Alternative activations such as hyperbolic tangent or sine produced comparable results, whereas ReLU was avoided because of its vanishing second derivative.

Network parameters were optimized using the Adam optimizer with an initial learning rate of 5×10−45 \times10^{-4} and an exponential decay (factor 0.1 every 1200 epochs). Training typically converged after approximately 4000 epochs, when the total loss L\mathcal{L} plateaued below 7×10−57 \times10^{-5}. The loss hyperparameters were λ1=λ2=λ3=λ4=1.0\lambda_1 = \lambda_2 = \lambda_3 = \lambda_4 = 1.0 and λ5=5.0\lambda_5 = 5.0. All hyperparameters—network depth/width, activation, and loss weights—were fixed based on preliminary experiments and standard PINN practices for fluid-dynamics problems [85]. Batch size had minimal impact on performance; we used a batch size of 256. All losses were computed using the mean-squared-error norm.

Collocation points for the PDE residual were uniformly sampled within a cubic domain enclosing the particle. At each training step, a random subset of NPDE≈5500N_{\mathrm{PDE}} \approx5500 points was used to evaluate LPDE\mathcal{L}_{\mathrm{PDE}}, while Ndata≈4300N_{\mathrm{data}} \approx4300 data points contributed to Ldata\mathcal{L}_{\mathrm{data}}, and NBC≈3721N_{\mathrm{BC}} \approx3721 boundary points enforced LBC\mathcal{L}_{\mathrm{BC}} on the particle surface. An additional Ntest≈29791N_{\mathrm{test}} \approx29791 points were used to evaluate predictive accuracy. Velocity data near the particle surface (r/a<1.2r/a < 1.2) were excluded to emulate experimental limitations. All quantities were nondimensionalized using the particle radius aa and characteristic velocity ∣U∣|U| to improve numerical conditioning.

The computational complexity of a forward–backward pass through a single network is O(LN2)\mathcal{O}(LN^2), where LL is the number of layers and NN the number of neurons per layer, corresponding to 9×1282≈1.5×1059 \times128^2 \approx1.5 \times10^5 operations per training point. Because the PDE residuals require first- and second-order derivatives via automatic differentiation, the effective cost increases by a constant factor. The cost per epoch, therefore, scales as O(8NtotalLN2)\mathcal{O}(8N_{\mathrm{total}}LN^2), where NtotalN_{\mathrm{total}} is the total number of PDE, data, and boundary points. With the parameters used here and the available GPU resources, each training epoch required approximately 4 seconds.

Algorithm 1 Training a PINN for active particle velocity and pressure fields

  1. Require:

    • Eight neural networks, each predicting one output: exterior flow uxoutu^{\mathrm{out}}_x, uyoutu^{\mathrm{out}}_y, uzoutu^{\mathrm{out}}_z, poutp^{\mathrm{out}} and surface/interface flow uxsu^s_x, uysu^s_y, uzsu^s_z, psp^s.

    • Training data: velocity values at training points.

    • Collocation points for PDE residual evaluation.

    • Boundary condition points.

    • Learning rate, batch size, and number of epochs.

    • Loss weights λi\lambda_i.

  2. Pre-processing: prepare training data and batch collocation points.

  3. Create neural networks with 9 hidden layers and 128 neurons per layer.

  4. Initialize optimizer and learning-rate scheduler.

  5. for epoch=1,…,epochtotalepoch = 1, \ldots, epoch_{\mathrm{total}} do

  6. Initialize losses: loss=0loss = 0.

  7. for each batch (x,y,z)(x, y, z) from PDE points do

  8. Zero gradients of network parameters WW and bb.

  9. Predict velocity and pressure fields.

  10. Compute total loss: loss=L=∑iλiLiloss = \mathcal{L} = \sum_i \lambda_i \mathcal{L}_i.

  11. Compute gradients: loss.backward().

  12. Update WW and bb using optimizer.step().

  13. Accumulate batch losses.

  14. End For

  15. Record total loss for this epoch.

  16. Update the learning rate using the scheduler.

  17. Save model weights and loss values for diagnostics.

  18. End For

  19. Save final trained parameters WW and $b

  20. Prediction: Apply the trained model on test data points to predict the velocity.

References

  1. [1]L. Scriven and C. Sterlning, The marangoni effects, Nature 187, 186 (1960).DOI
  2. [2]S. Vogel, Comparative biomechanics: life’s physical world (Princeton University Press, 2013).
  3. [3]E. Lauga and T. R. Powers, The hydrodynamics of swimming microorganisms, Rep. Prog. Phys. 72, 096601 (2009).DOI
  4. [4]M. C. Marchetti, J. F. Joanny, S. Ramaswamy, T. B. Liverpool, J. Prost, M. Rao, and R. A. Simha, Hydrodynamics of soft active matter, Rev. Mod. Phys. 85, 1143 (2013).arxiv.org/abs/1207.2929
  5. [5]K. Drescher, K. C. Leptos, I. Tuval, T. J. Pedley, and R. E. Goldstein, Dancing volvox: hydrodynamic bound states of swimming algae, Phys. Rev. Lett. 102, 168101 (2009).
  6. [6]J. Li, M. E. Krause, X. Chen, Y. Cheng, W. Dai, J. J. Hill, M. Huang, S. Jordan, D. LaCasse, L. Narhi, et al., Interfacial stress in the development of biologics: fundamental understanding, current practice, and future perspective, AAPS J. 21, 44 (2019).DOI
  7. [7]T. Toyota, N. Maru, M. M. Hanczyc, T. Ikegami, and T. Sugawara, Self-propelled oil droplets consuming “fuel” surfactant, J. Am. Chem. Soc. 131, 5012 (2009).
  8. [8]C. A. Weber, D. Zwicker, F. Jülicher, and C. F. Lee, Physics of active emulsions, Rep. Prog. Phys. 82, 064601 (2019).DOI
  9. [9]W. F. Paxton, K. C. Kistler, C. C. Olmeda, A. Sen, S. K. St. Angelo, Y. Cao, T. E. Mallouk, P. E. Lammert, and V. H. Crespi, Catalytic nanomotors: Autonomous movement of striped nanorods, J. Am. Chem. Soc. 126, 13424 (2004).pubmed.ncbi.nlm.nih.gov/15479099
  10. [10]Z. Izri, M. N. Van Der Linden, S. Michelin, and O. Dauchot, Self-propulsion of pure water droplets by spontaneous marangoni-stress-driven motion, Phys. Rev. Lett. 113, 248302 (2014).DOI
  11. [11]J. Palacci, S. Sacanna, S. H. Kim, G. R. Yi, D. J. Pine, and P. M. Chaikin, Light-activated self-propelled colloids, Philos. Trans. R. Soc. A 372, 20130372 (2014).DOI
  12. [12]C. C. Maass, C. Krüger, S. Herminghaus, and C. Bahr, Swimming droplets, Annu. Rev. Condens. Matter Phys. 7, 171 (2016).
  13. [13]C. H. Meredith, A. C. Castonguay, Y. J. Chiu, A. M. Brooks, P. G. Moerman, P. Torab, P. K. Wong, A. Sen, D. Velegol, and L. D. Zarzar, Chemical design of self-propelled Janus droplets, Matter 5, 616 (2022).DOI
  14. [14]K. Suzuno, D. Ueyama, M. Branicki, R. Tóth, A. Braun, and I. Lagzi, Maze solving using fatty acid chemistry, Langmuir 30, 9251 (2014).DOI
  15. [15]Z. Shen, A. Würger, and J. S. Lintuvuori, Hydrodynamic interaction of a self-propelling particle with a wall: Comparison between an active Janus particle and a squirmer model, Eur. Phys. J. E 41, 39 (2018).DOI
  16. [16]Z. Zhou, P. M. Vlahovska, and M. J. Miksis, Drag and torque coefficients of a translating particle with slip at a gas-liquid interface, Phys. Rev. Fluids 10, 104007 (2025).arxiv.org/abs/2410.09956
  17. [17]W. Uspal, M. N. Popescu, S. Dietrich, and M. Tasinkevych, Self-propulsion of a catalytically active particle near a planar wall: From reflection to sliding and hovering, Soft Matter 11, 434 (2015).DOI
  18. [18][18*] P. Bayati, M. N. Popescu, W. E. Uspal, S. Dietrich, and A. Najafi, Dynamics near planar walls for various model self-phoretic particles, Soft matter 15, 5644 (2019).DOI
  19. [19]B. Wu-Zhang, D. A. Fedosov, and G. Gompper, Collective behavior of squirmers in thin films, Soft Matter 20, 5687 (2024).
  20. [20]S. Kawakami and P. M. Vlahovska, Migration and deformation of a droplet enclosing an active particle, J. Fluid Mech. 1007, A41 (2025).DOI
  21. [21]F. A. Overberg, G. Gompper, and D. A. Fedosov, Motion of microswimmers in cylindrical microchannels, Soft Matter 20, 3007 (2024).DOI
  22. [22]S. Song, A. Llopis-Lorente, A. F. Mason, L. K. Abdelmohsen, and J. C. van Hest, Confined motion: motility of active microparticles in cell-sized lipid vesicles, J. Am. Chem. Soc. 144, 13831 (2022).DOI
  23. [23]R. Dey, C. M. Buness, B. V. Hokmabad, C. Jin, and C. C. Maass, Oscillatory rheotaxis of artificial swimmers in microchannels, Nat. Commun. 13, 2952 (2022).DOI
  24. [24]A. Zöttl and H. Stark, Hydrodynamics determines collective motion and phase behavior of active colloids in quasi-two-dimensional confinement, Phys. Rev. Lett. 112, 118101 (2014).pubmed.ncbi.nlm.nih.gov/24702421
  25. [25]P. Bayati and A. Najafi, Dynamics of two interacting active Janus particles, J. Chem. Phys. 144, 134901 (2016).DOI
  26. [26]S. Mallory, F. Alarcon, A. Cacciuto, and C. Valeriani, Self-assembly of active amphiphilic Janus particles, New J. Phys. 19, 125014 (2017).DOI
  27. [27]J. Clopés, G. Gompper, and R. G. Winkler, Alignment and propulsion of squirmer pusher–puller dumbbells, J. Chem. Phys. 156 (2022).DOI
  28. [28]G. Tucci, R. Golestanian, and S. Saha, Nonreciprocal collective dynamics in a mixture of phoretic Janus colloids, New J. Phys. 26, 073006 (2024).arxiv.org/abs/2402.09279
  29. [29]Y. Liu, R. Kailasham, P. G. Moerman, A. S. Khair, and L. D. Zarzar, Self-organized patterns in non-reciprocal active droplet systems, Angew. Chem. Int. Ed. 63, e202409382 (2024).DOI
  30. [30]D. Nie, Y. Ying, G. Guan, J. Lin, and Z. Ouyang, Two-dimensional study on the motion and interactions of squirmers under gravity in a vertical channel, J. Fluid Mech. 960, A31 (2023).DOI
  31. [31]P. Bayati and A. Najafi, Electrophoresis of active Janus particles, J. Chem. Phys. 150, 234902 (2019).DOI
  32. [32]S. Auschra, A. Bregulla, K. Kroy, and F. Cichos, Thermotaxis of Janus particles, Eur. Phys. J. E 44, 90 (2021).DOI
  33. [33]T. Ban, K. Tani, H. Nakata, and Y. Okano, Self-propelled droplets for extracting rare-earth metal ions, Soft Matter 10, 6316 (2014).DOI
  34. [34]C. Lozon, A. Cornet, S. Reculusa, P. Garrigue, A. Kuhn, and G. Salinas, Chemically-driven autonomous Janus electromagnets as magnetotactic swimmers, Angew. Chem. 136, e202408198 (2024).DOI
  35. [35]S. C. Takatori, R. De Dier, J. Vermant, and J. F. Brady, Acoustic trapping of active matter, Nat. Commun. 7, 10694 (2016).DOI
  36. [36]Z. Xiao, A. Name la, B. Garlan, and J. Simmchen, A platform for stop-flow gradient generation to investigate chemotaxis, Angew. Chem. 61, e202117768 (2022).DOI
  37. [37]P. Bayati and S. A. Mallory, Orbits, spirals, and trapped states: Dynamics of a phoretic Janus particle in a radial concentration gradient, ACS Nano 18, 23047 (2024).DOI
  38. [38][38*] V. E. Deneke, A. Puliafito, D. Krueger, A. V. Narla, A. De Simone, L. Primo, M. Vergassola, S. De Renzis, and S. Di Talia, Self-organized nuclear positioning syn- chronizes the cell cycle in drosophila embryos, Cell 177, 925 (2019).
  39. [39]P. H. Htet and E. Lauga, Cortex-driven cytoplasmic flows in elongated cells: fluid mechanics and application to nuclear transport in drosophila embryos, J. R. Soc. Interface 20, 20230428 (2023).arxiv.org/abs/2310.12586
  40. [40]P. H. Htet and E. Lauga, Analytical methods for cytoplasmic streaming in elongated cells, PNAS nexus 4, pgaf057 (2025).arxiv.org/abs/2502.10544
  41. [41]R. Theunissen, F. Scarano, and M. Riethmuller, On improvement of PIV image interrogation near stationary interfaces, Exp. Fluids 45, 557 (2008).DOI
  42. [42]C. J. Kähler, S. Scharnowski, and C. Cierpka, On the uncertainty of digital PIV and PTV near walls, Exp. Fluids 52, 1641 (2012).DOI
  43. [43]J. Katuri, W. E. Uspal, M. N. Popescu, and S. Sánchez, Inferring non-equilibrium interactions from tracer response near confined active Janus particles, Sci. Adv. 7, eabd0719 (2021).DOI
  44. [44]K. Singh, A. Yadav, P. Dwivedi, and R. Mangal, Interaction of active Janus colloids with tracers, Langmuir 38, 2686 (2022).DOI
  45. [45]R. W. Style, R. Boltyanskiy, G. K. German, C. Hyland, C. W. MacMinn, A. F. Mertz, L. A. Wilen, Y. Xu, and E. R. Dufresne, Traction force microscopy in physics and biology, Soft Matter 10, 4047 (2014).DOI
  46. [46]H. J. Butt, B. Cappella, and M. Kappl, Force measurements with the atomic force microscope: Technique, interpretation and applications, Surf. Sci. Rep. 59, 1 (2005).
  47. [47]W. J. Yan and Y. C. Ma, Shape reconstruction of an inverse stokes problem, J. Comput. Appl. Math. 216, 554 (2008).DOI
  48. [48]M. Raissi, P. Perdikaris, and G. E. Karniadakis, Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, J. Comput. Phys. 378, 686 (2019).DOI
  49. [49]G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang, Physics-informed machine learning, Nat. Rev. Phys. 3, 422 (2021).
  50. [50]S. Cai, Z. Mao, Z. Wang, M. Yin, and G. E. Karniadakis, Physics-informed neural networks (PINNs) for fluid mechanics: a review, Acta Mech. Sin. 37, 1727 (2021).arxiv.org/abs/2105.09506
  51. [51]C. Zhao, F. Zhang, W. Lou, X. Wang, and J. Yang, A comprehensive review of advances in physics-informed neural networks and their applications in complex fluid dynamics, Phys. Fluids 36, 101301 (2024).
  52. [52]M. Raissi, A. Yazdani, and G. E. Karniadakis, Hidden fluid mechanics: Learning velocity and pressure fields from flow visualizations, Science 367, 1026 (2020).DOI
  53. [53]A. Arzani, J. X. Wang, and R. M. D’Souza, Uncovering near-wall blood flow from sparse data with physics-informed neural networks, Phys. Fluids 33, 071905 (2021).DOI
  54. [54]X. Jin, S. Cai, H. Li, and G. E. Karniadakis, NSFnets (navier-stokes flow nets): Physics-informed neural networks for the incompressible navier-stokes equations, J. Comput. Phys. 426, 109951 (2021).DOI
  55. [55]D. R. Gurevich, M. R. Golden, P. A. Reinbold, and R. O. Grigoriev, Learning fluid physics from highly turbulent data using sparse physics-informed discovery of empirical relations (SPIDER), J. Fluid Mech. 996, A25 (2024).DOI
  56. [56]G. Kissas, Y. Yang, E. Hwuang, W. R. Witschey, J. A. Detre, and P. Perdikaris, Machine learning in cardiovascular flows modeling: Predicting arterial blood pressure from non-invasive 4D flow MRI data using physics-informed neural networks, Comput. Method Appl. Mech. Eng. 358, 12623 (2020).
  57. [57]M. Liu, L. Liang, and W. Sun, A generic physics-informed neural network-based constitutive model for soft biological tissues, Comput. Method Appl. Mech. Eng. 372, 113402 (2020).DOI
  58. [58]O. S. Pak and E. Lauga, Generalized squirming motion of a sphere, J. Eng. Math. 88, 1 (2014).arxiv.org/abs/1402.1070
  59. [59]M. Schmitt and H. Stark, Marangoni flow at droplet interfaces: Three-dimensional solution and applications, Physics of Fluids 28 (2016).DOI
  60. [60]C. Pozrikidis, A Practical Guide to Boundary Element Methods With the Software Library BEMLIB (CRC Press, 2002).
  61. [61]J. L. Moran and J. D. Posner, Phoretic self-propulsion, Annu. Rev. Fluid Mech. 49, 511 (2017).
  62. [62]S. Maiti, O. E. Shklyaev, A. C. Balazs, and A. Sen, Self-organization of fluids in a multienzymatic pump system, Langmuir 35, 3724 (2019).DOI
  63. [63]P. Schattling, B. Thingholm, and B. Stadler, Enhanced diffusion of glucose-fueled Janus particles, Chem. Mater. 27, 7412 (2015).DOI
  64. [64]X. Ma, A. Jannasch, U. R. Albrecht, K. Hahn, A. Miguel-López, E. Schaffer, and S. Sánchez, Enzyme-powered hollow mesoporous Janus nanomotors, Nano Lett. 15, 7043 (2015).DOI
  65. [65]K. D. Seo, J. Doh, and D. S. Kim, One-step microfluidic synthesis of Janus microhydrogels with anisotropic thermo-responsive behavior and organophilic/hydrophilic loading capability, Langmuir 29, 15137 (2013).DOI
  66. [66]F. Mou, C. Chen, Q. Zhong, Y. Yin, H. Ma, and J. Guan, Autonomous motion and temperature-controlled drug delivery of Mg/Pt-poly (N-Isopropylacrylamide) Janus micromotors driven by simulated body fluid and blood plasma, ACS Appl. Mater. Interfaces 6, 9897 (2014).DOI
  67. [67]R. Dong, Q. Zhang, W. Gao, A. Pei, and B. Ren, Highly efficient light-driven TiO2–Au Janus micromotors, ACS Nano 10, 839 (2016).DOI
  68. [68]J. L. Anderson, Colloid transport by interfacial forces, Annu. Rev. Fluid Mech. 21, 61 (1989).
  69. [69]B. V. Derjaguin, G. Sidorenkov, E. Zubashchenko, and E. Kiseleva, Kinetic phenomena in the boundary layers of liquids 1. the capillary osmosis, Prog. Surf. Sci. 43, 138 (1993).DOI
  70. [70]F. Cichos, K. Gustavsson, B. Mehlig, and G. Volpe, Machine learning for active matter, Nat. Mach. Intell. 2, 94 (2020).
  71. [71]J. Colen, M. Han, R. Zhang, S. A. Redford, L. M. Lemma, L. Morgan, P. V. Ruijgrok, R. Adkins, Z. Bryant, Z. Dogic, et al., Machine learning active-nematic hydrodynamics, Proc. Natl. Acad. Sci. U.S.A. 118, e2016708118 (2021).DOI
  72. [72]H. Hayano, A. Furukawa, and K. Kim, Learning microswimmer collision dynamics and predicting diffusivities using a neural-network-assisted boltzmann approach, arXiv preprint arXiv:2510.02559 (2025).
  73. [73]A. Shrestha, E. Kirkinis, and M. Olvera de la Cruz, Self-generated electrokinetic flows from active-charged boundary patterns, Phys. Rev. Res. 7, 023223 (2025).arxiv.org/abs/2412.15397
  74. [74]A. Merdasi, S. Ebrahimi, X. Yang, and R. Kunz, Physics informed neural network application on mixing and heat transfer in combined electroosmotic-pressure driven flow, Chem. Eng. Process. Process Intensif. 193, 109540 (2023).
  75. [75]S. Thakur, M. Raissi, and A. M. Ardekani, Viscoelastic-net: A physics informed neural network framework for stress discovery and model selection, J. Non-Newtonian Fluid Mech. 330, 105265 (2024).
  76. [76]B. Reyes, A. A. Howard, P. Perd ikaris, and A. M. Tartakovsky, Learning unknown physics of non-Newtonian fluids, Phys. Rev. Fluids 6 (2021).DOI
  77. [77]Y. Nakamura, S. Shiratori, R. Takagi, M. Sutoh, I. Sugihara, H. Nagano, and K. Shimano, Physics-informed neural network applied to surface-tension-driven liquid film flows, Int. J. Numer. Methods Fluids 94, 1359 (2022).DOI
  78. [78]M. Movahhedi, X. Y. Liu, B. Geng, C. Elemans, Q. Xue, J. X. Wang, and X. Zheng, Predicting 3D soft tissue dynamics from 2D imaging using physics informed neural networks, Commun. Biol. 6, 541 (2023).
  79. [79]G. Fu, S. Yuan, Y. Li, B. Sheng, Y. Cen, and Y. Lu, Physics-Informed Neural Network-based prediction of permeation performance in reverse osmosis membrane elements, Environ. Sci.: Water Res. Technol. 11, 3193 (2025).DOI
  80. [80]C. He, X. Hu, and L. Mu, A mesh-free method using piecewise deep neural network for elliptic interface problems, J. Comput. Appl. Math. 412, 114358 (2022).arxiv.org/abs/2005.04847
  81. [81]S. Wu and B. Lu, Inn: Interfaced neural networks as an accessible meshless approach for solving interface PDE problems, J. Comput. Phys. 470, 111588 (2022).DOI
  82. [82]H. Li, H. Fan, and Z. Tan, Two novel discontinuity-removing PINNs for solving variable coefficient elliptic interface problems on curved surfaces, Comput. Method Appl. Mech. Eng. 435, 117637 (2025).DOI
  83. [83]R. Qiu, R. Huang, Y. Xiao, J. Wang, Z. Zhang, J. Yue, Z. Zeng, and Y. Wang, Physics-informed neural networks for phase-field method in two-phase flow, Phys. Fluids 34, 052109 (2022).
  84. [84]A. K. Sarma, S. Roy, C. Annavarapu, P. Roy, and S. Jagannathan, Interface PINNs (I-PINNs): A physics-informed neural networks framework for interface problems, Comput. Method Appl. Mech. Eng. 429, 117135 (2024).DOI
  85. [85]S. Wang, S. Sankaran, H. Wang, and P. Perdikaris, An expert’s guide to training physics-informed neural networks, arXiv preprint arXiv:2308.08468 (2023).arxiv.org/abs/2308.08468

Paper details

vlnc:W_s338ug9z
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents