Deep Learning Surrogate for Fast CIR Prediction in Reactive Molecular Diffusion–Advection Channels

Meysam Ghanbari, Mohammad Taghi Dabiri, Mazen Hasna, Senior Member, IEEE · 2025

Read Deep Learning Surrogate for Fast CIR Prediction in Reactive Molecular Diffusion–Advection Channels by Meysam Ghanbari, Mohammad Taghi Dabiri, Mazen Hasna, Senior Member, IEEE (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-18071.txt.

Introduction

Meysam Ghanbari and four coauthors speed up design work for molecular messaging, where chemicals carry signals through tiny channels. Predicting the pulse a receiver will see means solving hard flow and reaction equations, and each new setting costs a fresh heavy run. Their shortcut trains a neural network as a stand in.

A detailed physics model first made a large set of sample pulses across wide ranges of flow, reaction, and shape. The network learned to map those settings straight to the pulse. On fresh test channels, 90 percent came out with an error score below 0.15. Designers can now scan options quickly and save the heavy math for final checks.

Catalog record

CREATORMeysam Ghanbari, Mohammad Taghi Dabiri, Mazen Hasna, Senior Member, IEEE
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE29 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-18071

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Deep Learning Surrogate for Fast CIR Prediction in Reactive Molecular Diffusion–Advection Channels Thanks: M. Ghanbari, M.T. Dabiri, and Tanvir Alam are with the College of Science and Engineering, Hamad Bin Khalifa University, Doha, Qatar (Email: [email protected]; [email protected]; [email protected]). Thanks: Mazen Hasna is with the Department of Electrical Engineering, Qatar University, Doha, Qatar (Email: [email protected]). Thanks: Khalid A. Qaraqe is with the College of Science and Engineering, Hamad Bin Khalifa University, Doha, Qatar, and also with the Department of Electrical Engineering, Texas A&M University at Qatar, Doha, Qatar (Email: [email protected]). Thanks: This publication was made possible by Texas A&M University at Qatar and Hamad Bin Khalifa University, which supported this publication. Meysam Ghanbari, Mohammad Taghi Dabiri, Mazen Hasna, Senior Member, IEEE Abstract Accurate channel impulse response (CIR) modeling in molecular communication (MC) often requires solving coupled reactive diffusion–advection equations, which is computationally expensive for large parameter sweeps or design loops. We develop a deep-learning surrogate for a three-dimensional duct MC channel with reactive diffusion–advection transport and reversible ligand–receptor binding on a finite ring receiver. Using a physics-based partial differential equation (PDE)–ordinary differential equation (ODE) model, we generate a large CIR dataset across broad transport, reaction, and geometric ranges and train a neural network that maps these parameters directly to the CIR. On an independent test set, the surrogate closely matches reference CIRs both qualitatively and quantitatively: the empirical cumulative distribution function (CDF) of the normalized root mean square error (NRMSE) shows that 90% of test channels are predicted with error below 0.15, with only weak dependence on individual parameters. The surrogate therefore offers an accurate and computationally efficient replacement for repeated PDE-based CIR evaluations in MC system analysis and design. I introduction Molecular communication (MC) is an emerging paradigm for nano–bio interfaces and microfluidic platforms, where information is carried by signaling molecules rather than electromagnetic waves. In a typical MC link, a transmitter releases molecules into a medium where they diffuse, experience flow, and may degrade before reaching a receiver with reactive or receptor-coated surfaces [ 5 ] . These transport and reaction processes determine how the transmitted signal is distorted. The channel impulse response (CIR) captures their combined effect and underpins modulation, detection, synchronization, and performance analysis. Accurate and efficient CIR prediction in realistic geometries is therefore crucial for practical MC system design. However, for most realistic MC channels, no analytical closed-form expression for the CIR is available. Instead, the CIR must usually be computed by numerically solving a coupled 3D advection–diffusion–reaction partial differential equation (PDE) together with surface ordinary differential equations (ODE) that describe receptor binding and unbinding at the receiver. These models are often stiff and computationally demanding, particularly when many CIR evaluations are required for parameter sweeps, system optimization, or adaptive link operation in real time [ 6 ] . The authors in [ 2 ] studied heterogeneity and mobility in an unbounded 3D purely diffusive molecular channel, where the diffusivity of nodes and molecules varies randomly in time, leading to a non-Gaussian, time-varying CIR. In [ 8 ] , a bounded but still diffusion-driven spherical shell is considered, and an analytical CIR is obtained for partially absorbing inner and outer surfaces with optional degradation. Multiple absorbing receivers are handled in [ 4 ] by modeling diffusion toward several fully absorbing spheres in an unbounded medium and expressing the problem as coupled integral equations. A more general bounded model with mixed absorbing and reflecting patches on circular and spherical boundaries is presented in [ 3 ] , where the heterogeneous surface is homogenized into an equivalent partially reactive boundary under diffusion-only transport. On the received-signal side, [ 1 ] considers a spherical transmitter–receiver pair with partial absorption in an unbounded diffusive channel and fits a low-dimensional analytical CIR model to simulations. Finally, [ 14 ] is based on a 1D diffusion channel with a Poisson reception process and uses this setting to study identification-style communication. For duct-like and intrabody scenarios, [ 13 ] models micro-circulation for bio-Internet-of-Things applications using a laminar Poiseuille flow network of arterioles, capillaries, and venules, where large dimensionless numbers justify advection-dominated transport and flow-driven CIRs are derived for each vessel segment, without explicitly modeling localized reactive receiver patches. Data-driven approaches complement these physical models. In [ 12 ] , neural detectors are trained on synthetic CIRs and experimental traces, and explainability tools indicate that the learned networks effectively implement intuitive decision rules similar to classical peak- and edge-based detectors, while their behavior can still be interpreted using standard signal-to-noise–based BER concepts. In [ 9 ] , the aim is to approximate the CIR itself for complex 2D diffusion-based topologies: transmitter and receiver layouts and reflecting obstacles are encoded as images and passed to a convolutional and fully connected neural network, which predicts the CIR shape and peak and can be superposed to approximate multi-transmitter channels with low normalized error compared to Monte-Carlo simulations. Overall, these works demonstrate substantial progress in modeling molecular channels, but they also rely on simplifying assumptions. Analytical studies often use unbounded or highly idealized geometries and omit either flow or detailed surface reactions, while data-driven surrogates are trained on these simplified channels or focus on detection rather than physically faithful CIR prediction. These assumptions make the problems tractable, but in realistic ducts with laminar flow and reactive surfaces (e.g., in blood vessels or microfluidic channels) they can cause mismatches between the predicted channel behavior and the underlying transport dynamics and require recomputing the channel response whenever the physical parameters change. In this work, we develop a 3D duct model that jointly captures diffusion, laminar flow, and surface reaction at a finite ring-shaped wall receiver and use it to generate a large library of CIRs over a wide range of realistic channel conditions. On top of this model, we train a deep learning surrogate that maps physical parameters of the duct, the flow, and the surface chemistry directly to the full temporal channel response at low computational cost. The resulting framework serves as a benchmark channel model for evaluating detectors, equalizers, and coding schemes for reactive diffusion–advection links, and as a “ground-truth” emulator against which future analytical approximations and data-driven surrogates can be compared. II system model We consider a MC link inside a cylindrical duct as illustrated in Fig. 1 , where signaling molecules propagate under diffusion, background flow, and first-order degradation. The transmitter is located at the center of the duct’s cross-section, and molecules spread throughout the fluid as they travel downstream. The duct wall is reflective everywhere except for a narrow ring-shaped sensing region that serves as the receiver. Molecules that reach this ring can bind to surface receptors and generate the measurable output signal, while all other molecules either continue propagating or undergo degradation within the channel. II-A Transmitter We consider a finite-volume transmitter positioned on the duct axis at \mathbf{r}_{\mathrm{tx}}=(0,0,z_{\mathrm{tx}}) , which releases N_{0} identical signaling molecules of type A at time t=0 . The transmitter occupies a small spherical region of radius a_{\mathrm{tx}}\ll a_{c} , where a_{c} is the duct radius. At the moment of release, the molecules are assumed to be uniformly distributed within the transmitter volume, leading to the following initial concentration field [ 6 ] : c(\mathbf{r},0)=\frac{N_{0}}{V_{\mathrm{tx}}}\,\Pi\!\left(\frac{\lVert\mathbf{r}-\mathbf{r}_{\mathrm{tx}}\rVert}{a_{\mathrm{tx}}}\right), (1) where \Pi(\cdot) is

Read the complete text