Decoding the Neural Enigma: Digital Twins of Neurons Revolutionize Brain Research

For years, scientists have relied on traditional cell culture experiments to study the intricate functions of neurons within the brain. Understanding these interactions is key to addressing major neurodevelopmental disorders such as autism spectrum disorder and attention-deficit/hyperactivity disorder (ADHD), as well as neurodegenerative diseases such as Alzheimer’s, Parkinson’s, and amyotrophic lateral sclerosis (ALS). Despite being the industry standard, conventional experimentation methods are not only time-intensive and burdensome, but also severely limited in the variables and scale they can explore. These limitations have significantly slowed our pace in finding effective treatments for neurodevelopmental and neurodegenerative conditions.

To address this challenge, the scientific community welcomes a profound paradigm shift with the creation of “digital twins” of neurons. These highly realistic virtual models combine advanced computational models with deep learning technologies to provide researchers with unprecedented levels of efficiency at a low cost, allowing them to simulate neuronal growth, development, and organelle transport at speeds and scales previously thought unimaginable. Nevertheless, fundamental challenges with mathematical modeling and computational efficiency remain. We focus on two key aspects of the digital twin paradigm: physics-based foundational modeling and the speed revolution brought by artificial intelligence (AI) surrogate models.

Physics-Based Modeling and Simulation Challenge

To build digital models capable of accurately predicting neuronal behavior, we must first return to our roots: describing the complex dynamics of neurons through precise mathematical and fundamental physical principles. These techniques are not only the cornerstone of understanding neuronal function, but also the solid foundation upon which all subsequent AI acceleration technologies are built.

One such physics-based model—designed to simulate the neuronal growth process—skillfully combines the phase field method with isogeometric analysis (IGA) [7, 9, 10]. The key innovation of this method is that it is not purely theoretical calculation; instead, it deeply integrates experimental data from hippocampal neurons of rats. Mathematically, the morphological evolution of the neuron is described by the governing equation of the phase field \(\phi\), which captures anisotropy and driving forces:

\[\frac{\partial\phi}{\partial{t}} = M_{\phi} \bigg[\nabla\;\dot\;\;\bigg(a(\Psi)^2\nabla\phi-\frac{\partial}{\partial{x}}\big(a(\Psi)\frac{\partial{a}(\Psi)}{\partial\Psi}\frac{\partial\phi}{\partial{y}}\bigg)+\]

\[\frac{\partial}{\partial{y}}\bigg(a(\Psi)\frac{\partial{a}(\Psi)}{\partial\Psi}\frac{\partial\phi}{\partial{x}}\bigg)+\phi(1-\phi)(\phi-0.5+F_{driv}+6H|\nabla\phi|)\bigg]\]

where \(a(\Psi)\) represents the anisotropy coefficient and \(F_{driv}\) is the driving force for neuronal growth. Real data serves as driving and constraining conditions, guiding the growth of the simulated neuron to exhibit highly biomimetic patterns (see Figures 1a and 1b).

Simultaneously, the interior of a neuron acts as a bustling material transport network. To deeply investigate this mechanism, we utilized IGA solvers to establish a motor-assisted transport model within complex three-dimensional neuronal tree structures [4, 5] (see Figure 1c). This model describes the partial differential equation-constrained optimization system by coupling the Navier-Stokes equations with advection-diffusion-reaction equations. The dynamic change in material concentration is governed by minimizing

\[\mathcal{J}(n_{\pm},v_{\pm},f_{\pm})=\frac{1}{2}\int^T_0\int_{\Omega}(v_{\pm}-V_{\pm})\;d{\Omega}dt+\frac{\alpha}{2}\int^T_0\int_{\Omega}||\nabla{n_{\pm}}||^2d\;{\Omega}dt+\frac{\beta}{2}\int^T_0\int_{\Omega}f^2_{\pm}d{\Omega}dt \quad \textrm{in}\quad \Omega\times(0,T], \]

subject to 

\[\frac{\partial{n_0}}{\partial{t}}=-(k_+-k_-)n_0+k'_+l_+n_++k'_-l_-n_- \; \textrm{in}\;\Omega\times(0,T],\]\[\frac{\partial{l_{\pm}}{n_{\pm}}}{\partial{t}}=-v_{\pm}\;\cdot\;\nabla(l_{\pm}n_{\pm})\pm\; D_{\pm}\nabla^2(l_{\pm}n_{\pm})+ \;k_{\pm}n_0-k'_{\pm}(l_{\pm}n_{\pm})\;\textrm{in}\;\Omega\times(0,T],\]\[\frac{\partial{v_{\pm}}}{\partial{t}}+v_{\pm}\;\cdot\;\nabla{v_{\pm}}=-\nabla{n_{\pm}}+\nabla\;\cdot\;({\mu}\nabla{v_{\pm}})+f_{\pm}\; \textrm{in}\;\Omega\times(0,T].\]

Here, \(n_0\) is the concentration of free material, while \(n_+\) represents the concentration of material bound to microtubules that are moving either forward or backward. The definition of other parameters can be found in [4, 5]. This high-fidelity simulation can meticulously reproduce material flow inside neurons and successfully simulate the “traffic jam” phenomena caused by microtubule structural anomalies (see Figure 1d).

However, despite the immense power of these physics-based simulation methods, their high computational cost has become a major bottleneck for widespread research adoption.

Novel Approaches to Surrogate Modeling

To break through the high cost barrier, the fields of neuroscience and computational science are undergoing an AI-driven revolution. AI surrogate models provide the critical pillars for a digital twin to mirror the full life cycle of a living neuron, from construction and operation to deterioration.

Convolutional neural networks (CNNs) show amazing potential for predicting neuronal growth by utilizing specialized machine learning (ML) techniques that can mimic the human cortex. We established a CNN-based surrogate model centered on a customized convolutional autoencoder [6]. Compared to traditional IGA physics-based simulators, our CNN model reduces computation time by a staggering seven orders of magnitude while maintaining an accuracy of \(97.77\) percent when predicting complex neurite growth patterns.

When addressing neurological disorders, understanding neuronal deterioration is crucial. Recent research introduces a high-throughput ML framework that utilizes a MetaFormer architecture combined with a gated spatiotemporal attention mechanism to precisely predict neurite deterioration patterns [8] (see Figure 2a). This framework integrates synthetic data generated by IGA phase-field models with experimental imagery, allowing for the capture of long-range temporal dependencies and intricate morphological transformations. This technology achieved low average error rates—\(1.96\) percent on synthetic data and \(6.03\) percent on experimental data—successfully overcoming the challenge of scarce experimental datasets and providing a powerful predictive tool for studying neural disorders.

In the complex realm of material transport simulation, Graph-Autoencoder-based Latent Dynamics Surrogate (GALDS) has emerged as another innovative surrogate model [1]. GALDS employs an advanced workflow operating in a concise latent space, using neural ordinary differential equations to efficiently predict the system’s dynamic evolution over time:

\[\frac{dz(t)}{dt}=f_0(z(t),t). \]

This “compress-predict-reconstruct” approach surpasses previous methods such as the physics-informed graph neural network (PGNN), which learns directly in complex physical spaces [3, 5]. By operating in this latent space, GALDS saves over \(99\) percent of computational resources while maintaining high accuracy (see Figure 2b).

Challenges and Opportunities

Starting from the monumental challenge of understanding brain function, we have witnessed a clear path of technological evolution: traditional physics-based simulations laid the foundation, while deep learning-driven AI surrogate models such as CNNs, MetaFormer, and GALDS thoroughly shattered efficiency bottlenecks.

These advancements do not merely make existing processes faster; they unlock a completely new realm of possibility for neuroscience research. Scientists are no longer limited to long experimental cycles that can only test one hypothesis at a time. Instead, they can conduct large-scale “digital scans” within the digital world: systematically testing the effects of thousands of genetic mutations, screening the responses of hundreds of potential drug compounds, or exploring vast parameter spaces to discover unexpected pathological mechanisms.

Looking ahead, as digital twin models of neurons mature and become more widespread, it will vastly accelerate our understanding of the root causes of neurological disorders. This technology offers a powerful computational microscope, allowing us to witness the onset and progression of disease at the cellular level, thus unlocking the ability to screen for potential therapeutic targets more rapidly. It represents not just a leap in computational power, but a beacon of hope, promising to illuminate the unknown corners of neuroscience and significantly shorten the long road to finding effective treatments for neurological diseases.

This article is based on Jessica Zhang’s AWM-SIAM Sonia Kovalevsky Prize Lecture at the Third Joint SIAM/CAIMS Annual Meetings which took place in Montréal, Québec, Canada.

Acknowledgements: The presented work was supported by the U.S. National Science Foundation grants CMMI-1953323 and CBET-2332084. This work used RM-nodes on Bridges-2 Supercomputer at Pittsburgh Supercomputer center through allocation ID eng170006p from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by the National Science Foundation, United States grants #2138259, #2138307, #2137603, and #2138296.

References 
[1] Hsieh, T.Y. & Zhang, Y.J. (2025). GALDS: A graph-autoencoder-based latent dynamics surrogate model to predict neurite material transport. Comput. Method Appl. Mech. Eng., 447, 118409.
[2] Li, A., Chai, X., Yang, G., & Zhang, Y.J. (2019). An Isogeometric analysis computational platform for material transport simulation in complex neurite networks. MCB Mol. Cell. Biomech., 16(2), 123-140.
[3] Li, A., Farimani, A.B., & Zhang, Y.J. (2021). Deep learning of material transport in complex neurite networks. Sci. Rep., 11, 11280.
[4] Li, A. & Zhang, Y.J. (2022). Modeling intracellular transport and traffic jam in 3D neurons using PDE-constrained optimization. J. Mech., 38, 44-59.
[5] Li, A. & Zhang, Y.J. (2023). Isogeometric analysis-based physics-informed graph neural network for studying traffic jam in neurons. Comput. Method Appl. Mech. Eng., 403, 115757.
[6] Qian, K., Liao, A.S., Gu, S., Webster-Wood, V.A., & Zhang, Y.J. (2023). Biomimetic IGA neuron growth modeling with neurite morphometric features and CNN-based prediction. Comput. Method Appl. Mech. Eng., 417, 116213.
[7] Qian, K., Pawar, A., Liao, A., Anitescu, C., Webster-Wood, V., Feinberg, A.W., ... Zhang, Y.J. (2022). Modeling neuron growth using isogeometric collocation based phase field method. Sci. Rep., 12, 8120.
[8] Qian, K., Suarez, G.O., Nambara, T., Kanekiyo, T., & Zhang, Y.J. (2025). High-throughput machine learning framework for predicting neurite deterioration using MetaFormer attention. Comput. Methods Appl. Mech. Eng., 442, 118003.
[9] Qian, K., Suarez, G.O., Nambara, T., Kanekiyo, T., Liao, A. S., Webster-Wood, V.A., & Zhang, Y.J. (2025). Neurodevelopmental disorders modeling using isogeometric analysis, dynamic domain expansion and local refinement. Comput. Method Appl. M., 433(B), 117534.
[10] Qian, K., & Zhang, Y.J. (2025). 3D neuron growth and neurodevelopmental disorder modeling based on truncated hierarchical B-Splines with multi-level local refinements. Comput. Method Appl. M., 445, 118176.